• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2019-2020 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/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.h"
25 
26 #include "arm_compute/core/CL/CLHelpers.h"
27 #include "arm_compute/core/CL/CLKernelLibrary.h"
28 #include "arm_compute/core/CL/ICLTensor.h"
29 #include "arm_compute/core/CL/OpenCL.h"
30 #include "arm_compute/core/Helpers.h"
31 #include "arm_compute/core/TensorInfo.h"
32 #include "arm_compute/core/Utils.h"
33 #include "arm_compute/core/Validate.h"
34 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
35 #include "src/core/AccessWindowStatic.h"
36 #include "src/core/helpers/AutoConfiguration.h"
37 #include "src/core/helpers/WindowHelpers.h"
38 #include "support/StringSupport.h"
39 
40 #include <cstddef>
41 #include <cstdint>
42 #include <tuple>
43 
44 using namespace arm_compute::misc::shape_calculator;
45 
46 namespace arm_compute
47 {
48 namespace
49 {
50 using ElementsProcessed = Steps;
51 
validate_arguments(const ITensorInfo * input0,const ITensorInfo * input1,const ITensorInfo * output,const GEMMKernelInfo & gemm_info,const ITensorInfo * vector_sum_col,const ITensorInfo * vector_sum_row,const ITensorInfo * bias,const ITensorInfo * output_multipliers,const ITensorInfo * output_shifts)52 Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, const GEMMKernelInfo &gemm_info,
53                           const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias,
54                           const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
55 {
56     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output);
57     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
58     if(input0->data_type() == DataType::QASYMM8)
59     {
60         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
61     }
62     else
63     {
64         ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::QASYMM8, DataType::QSYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8_PER_CHANNEL);
65     }
66     ARM_COMPUTE_RETURN_ERROR_ON_MSG(input0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4");
67     ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3");
68 
69     const GEMMRHSMatrixInfo       rhs_info     = gemm_info.rhs_info;
70     const GEMMLHSMatrixInfo       lhs_info     = gemm_info.lhs_info;
71     const GEMMLowpOutputStageInfo output_stage = gemm_info.output_stage;
72 
73     ARM_COMPUTE_RETURN_ERROR_ON_MSG((((rhs_info.k0 & (rhs_info.k0 - 1)) && rhs_info.k0 != 3) || (rhs_info.k0 > 16)), "Only 2,3,4,8,16 are supported for k0");
74     ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 1 || lhs_info.m0 > 8);
75     ARM_COMPUTE_RETURN_ERROR_ON_MSG((((rhs_info.n0 & (rhs_info.n0 - 1)) && rhs_info.n0 != 3) || rhs_info.n0 > 16), "Only 2,3,4,8,16 are supported for n0");
76     ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs_info.export_to_cl_image, "Export to CLImage not supported for quantized GEMM");
77 
78     const int m = gemm_info.m;
79     const int n = gemm_info.n;
80     const int k = gemm_info.k;
81 
82     TensorShape tensor_shape1{ input1->tensor_shape() };
83     tensor_shape1.set(0, n);
84     tensor_shape1.set(1, k);
85 
86     const TensorInfo tensor_info1          = input1->clone()->set_tensor_shape(tensor_shape1);
87     const TensorInfo tensor_info_reshaped1 = input1->clone()->set_tensor_shape(compute_rhs_reshaped_shape(tensor_info1, rhs_info));
88 
89     ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(0) != static_cast<unsigned int>(k));
90     if(gemm_info.reinterpret_input_as_3d)
91     {
92         ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(1) * input0->dimension(2) != static_cast<unsigned int>(m));
93     }
94     else
95     {
96         ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(1) != static_cast<unsigned int>(m));
97     }
98     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, &tensor_info_reshaped1);
99 
100     const TensorShape expected_output_shape = compute_mm_shape(*input0, *input1, gemm_info);
101     if(output->total_size() != 0)
102     {
103         const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(expected_output_shape);
104         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
105         if(output_stage.type == GEMMLowpOutputStageType::NONE)
106         {
107             ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
108         }
109         else
110         {
111             ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output);
112         }
113     }
114 
115     if(bias != nullptr)
116     {
117         ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32);
118         ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
119         ARM_COMPUTE_RETURN_ERROR_ON(expected_output_shape[0] != bias->dimension(0));
120     }
121 
122     ARM_COMPUTE_RETURN_ERROR_ON_MSG((output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN) || (output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FLOAT),
123                                     "Only GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT is supported");
124 
125     // Checks performed if the output stage needs to be fused
126     if(output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT)
127     {
128         // If a_offset == 0, vector_sum_col can be a nullptr
129         if(gemm_info.a_offset != 0)
130         {
131             ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_col, 1, DataType::S32);
132             ARM_COMPUTE_RETURN_ERROR_ON(vector_sum_col->dimension(0) != expected_output_shape[0]);
133         }
134 
135         // If b_offset == 0, vector_sum_row can be a nullptr
136         if(gemm_info.b_offset != 0)
137         {
138             ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_row, 1, DataType::S32);
139 
140             // Check if mm result is a 3D reinterpretation
141             const bool reinterpret_as_3d = expected_output_shape.num_dimensions() > 1 && expected_output_shape.y() != vector_sum_row->tensor_shape().x();
142 
143             // Validate input
144             ARM_COMPUTE_RETURN_ERROR_ON(reinterpret_as_3d && vector_sum_row->dimension(0) != (expected_output_shape[1] * expected_output_shape[2]));
145             ARM_COMPUTE_RETURN_ERROR_ON(!reinterpret_as_3d && vector_sum_row->dimension(0) != expected_output_shape[1]);
146 
147             if(expected_output_shape.num_dimensions() > 1)
148             {
149                 const unsigned int output_batch_idx = reinterpret_as_3d ? 3 : 2;
150 
151                 TensorShape vector_sum_row_shape = vector_sum_row->tensor_shape();
152                 vector_sum_row_shape.collapse_from(1);
153                 TensorShape collapsed_output_shape(expected_output_shape);
154                 collapsed_output_shape.collapse_from(output_batch_idx);
155 
156                 ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_row_shape[1] != collapsed_output_shape[output_batch_idx],
157                                                 "vector_sum_row must have the same number of batches of output tensor");
158 
159                 if(gemm_info.a_offset != 0)
160                 {
161                     TensorShape vector_sum_col_shape = vector_sum_col->tensor_shape();
162                     vector_sum_col_shape.collapse_from(1);
163 
164                     ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_col_shape[1] != 1 && vector_sum_col_shape[1] != vector_sum_row_shape[1],
165                                                     "vector_sum_col tensor must have the same number of batches of vector_sum_row_shape or the number of batches must be set to 1");
166                 }
167             }
168         }
169 
170         if(output->total_size() != 0)
171         {
172             ARM_COMPUTE_RETURN_ERROR_ON(output_stage.output_data_type != output->data_type());
173         }
174         ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_min_bound > output_stage.gemmlowp_max_bound);
175 
176         if(output_multipliers != nullptr && output_shifts != nullptr)
177         {
178             ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_multipliers, 1, DataType::S32);
179             ARM_COMPUTE_RETURN_ERROR_ON(output_multipliers->num_dimensions() > 1);
180             ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_shifts, 1, DataType::S32);
181             ARM_COMPUTE_RETURN_ERROR_ON(output_shifts->num_dimensions() > 1);
182             if(output_stage.is_quantized_per_channel)
183             {
184                 ARM_COMPUTE_RETURN_ERROR_ON(expected_output_shape[0] != output_shifts->dimension(0));
185                 ARM_COMPUTE_RETURN_ERROR_ON(expected_output_shape[0] != output_multipliers->dimension(0));
186             }
187         }
188     }
189     return Status{};
190 }
191 
validate_and_configure_window(ITensorInfo * input0,ITensorInfo * input1,ITensorInfo * output,const GEMMKernelInfo & gemm_info,ITensorInfo * vector_sum_col,ITensorInfo * vector_sum_row,ITensorInfo * bias,ITensorInfo * output_multipliers,ITensorInfo * output_shifts,ElementsProcessed & num_elements_processed)192 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output, const GEMMKernelInfo &gemm_info,
193                                                         ITensorInfo *vector_sum_col, ITensorInfo *vector_sum_row, ITensorInfo *bias,
194                                                         ITensorInfo *output_multipliers, ITensorInfo *output_shifts, ElementsProcessed &num_elements_processed)
195 {
196     const GEMMLowpOutputStageInfo output_stage = gemm_info.output_stage;
197 
198     unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
199     unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
200     bool          reinterpret_input_as_3d             = gemm_info.reinterpret_input_as_3d;
201     bool          reinterpret_output_as_3d            = (gemm_info.depth_output_gemm3d != 0);
202 
203     Window win{};
204     Window win_out{};
205     bool   window_changed = false;
206 
207     // In case both input and output have to be reinterpreted as 3D tensors,
208     // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
209     if(reinterpret_input_as_3d == reinterpret_output_as_3d)
210     {
211         reinterpret_output_as_3d = false;
212     }
213 
214     // Output tensor auto initialization if not yet initialized
215     const TensorShape expected_output_shape = compute_mm_shape(*input0, *input1, gemm_info);
216     if(output_stage.type != GEMMLowpOutputStageType::NONE)
217     {
218         auto_init_if_empty(*output, input0->clone()->set_tensor_shape(expected_output_shape).set_data_type(output_stage.output_data_type));
219     }
220     else
221     {
222         auto_init_if_empty(*output, input0->clone()->set_tensor_shape(expected_output_shape).set_data_type(DataType::S32));
223     }
224 
225     TensorInfo tmp_info(*output);
226 
227     if(reinterpret_output_as_3d)
228     {
229         // Since the output tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM,
230         // the window needs to be constructed on the 2D collapsed version of the tensor
231         TensorShape tmp_shape(output->tensor_shape());
232         tmp_shape.collapse(2U, 1U);
233         tmp_info.set_tensor_shape(tmp_shape);
234     }
235 
236     // Configure kernel window
237     num_elems_processed_per_iteration_x = gemm_info.rhs_info.n0;
238     num_elems_processed_per_iteration_y = gemm_info.lhs_info.m0;
239 
240     win     = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
241     win_out = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
242 
243     if(output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT)
244     {
245         if(gemm_info.a_offset != 0)
246         {
247             AccessWindowHorizontal vector_sum_col_access(vector_sum_col, 0, num_elems_processed_per_iteration_x);
248             window_changed = window_changed || update_window_and_padding(win_out, vector_sum_col_access);
249         }
250         // No access window needed for vector_sum_row
251         ARM_COMPUTE_UNUSED(vector_sum_row);
252 
253         if(bias != nullptr)
254         {
255             AccessWindowHorizontal bias_access(bias, 0, num_elems_processed_per_iteration_x);
256             window_changed = window_changed || update_window_and_padding(win_out, bias_access);
257         }
258 
259         if(output_multipliers != nullptr && output_multipliers->dimension(0) > 1)
260         {
261             AccessWindowHorizontal output_multipliers_access(output_multipliers, 0, num_elems_processed_per_iteration_x);
262             AccessWindowHorizontal output_shifts_access(output_shifts, 0, num_elems_processed_per_iteration_x);
263             window_changed = window_changed || update_window_and_padding(win_out, output_multipliers_access, output_shifts_access);
264         }
265     }
266 
267     // Collapse along the Z direction
268     // This collapse needs to be here in order to tune the Z dimension of LWS
269     Window             collapsed             = win;
270     const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(output->num_dimensions()), 2u);
271     collapsed                                = win.collapse(win, dimension_to_collapse);
272 
273     Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
274     return std::make_pair(err, collapsed);
275 }
276 } // namespace
277 
CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel()278 CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel()
279     : _input0(nullptr),
280       _input1(nullptr),
281       _output(nullptr),
282       _vector_sum_col(nullptr),
283       _vector_sum_row(nullptr),
284       _bias(nullptr),
285       _output_multipliers(nullptr),
286       _output_shifts(nullptr),
287       _slide_matrix_b(true),
288       _reinterpret_input_as_3d(false),
289       _reinterpret_output_as_3d(false),
290       _use_dummy_work_items(false),
291       _is_quantized_per_channel(false),
292       _fuse_output_stage(false)
293 {
294 }
295 
configure(const ICLTensor * input0,const ICLTensor * input1,ICLTensor * output,const GEMMKernelInfo & gemm_info,const ICLTensor * vector_sum_col,const ICLTensor * vector_sum_row,const ICLTensor * bias,const ICLTensor * output_multipliers,const ICLTensor * output_shifts)296 void CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, const GEMMKernelInfo &gemm_info,
297                                                               const ICLTensor *vector_sum_col, const ICLTensor *vector_sum_row, const ICLTensor *bias,
298                                                               const ICLTensor *output_multipliers, const ICLTensor *output_shifts)
299 {
300     configure(CLKernelLibrary::get().get_compile_context(), input0, input1, output, gemm_info, vector_sum_col, vector_sum_row, bias, output_multipliers, output_shifts);
301 }
302 
configure(const CLCompileContext & compile_context,const ICLTensor * input0,const ICLTensor * input1,ICLTensor * output,const GEMMKernelInfo & gemm_info,const ICLTensor * vector_sum_col,const ICLTensor * vector_sum_row,const ICLTensor * bias,const ICLTensor * output_multipliers,const ICLTensor * output_shifts)303 void CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output,
304                                                               const GEMMKernelInfo &gemm_info,
305                                                               const ICLTensor *vector_sum_col, const ICLTensor *vector_sum_row, const ICLTensor *bias,
306                                                               const ICLTensor *output_multipliers, const ICLTensor *output_shifts)
307 {
308     ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output);
309     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(),
310                                                   input1->info(),
311                                                   output->info(),
312                                                   gemm_info,
313                                                   vector_sum_col != nullptr ? vector_sum_col->info() : nullptr,
314                                                   vector_sum_row != nullptr ? vector_sum_row->info() : nullptr,
315                                                   bias != nullptr ? bias->info() : nullptr,
316                                                   output_multipliers != nullptr ? output_multipliers->info() : nullptr,
317                                                   output_shifts != nullptr ? output_shifts->info() : nullptr));
318 
319     auto                          padding_info = get_padding_info({ input0, input1, output, vector_sum_row });
320     const GEMMRHSMatrixInfo       rhs_info     = gemm_info.rhs_info;
321     const GEMMLHSMatrixInfo       lhs_info     = gemm_info.lhs_info;
322     const GEMMLowpOutputStageInfo output_stage = gemm_info.output_stage;
323     const int32_t                 a_offset     = gemm_info.a_offset;
324     const int32_t                 b_offset     = gemm_info.b_offset;
325 
326     _input0                   = input0;
327     _input1                   = input1;
328     _output                   = output;
329     _vector_sum_col           = vector_sum_col;
330     _vector_sum_row           = vector_sum_row;
331     _bias                     = bias;
332     _output_multipliers       = output_multipliers;
333     _output_shifts            = output_shifts;
334     _reinterpret_input_as_3d  = gemm_info.reinterpret_input_as_3d;
335     _reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d != 0);
336     _use_dummy_work_items     = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device());
337     _is_quantized_per_channel = output_stage.is_quantized_per_channel;
338 
339     // In case both input and output have to be reinterpreted as 3D tensors,
340     // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
341     if(_reinterpret_input_as_3d == _reinterpret_output_as_3d)
342     {
343         _reinterpret_input_as_3d  = false;
344         _reinterpret_output_as_3d = false;
345     }
346 
347     // Check if we need to slide the matrix B
348     const unsigned int num_dimensions_input0 = _input0->info()->num_dimensions();
349     _slide_matrix_b                          = (_input1->info()->num_dimensions() >= num_dimensions_input0);
350 
351     ElementsProcessed num_elements_processed{};
352 
353     // Configure kernel window
354     auto win_config = validate_and_configure_window(input0->info(),
355                                                     input1->info(),
356                                                     output->info(),
357                                                     gemm_info,
358                                                     vector_sum_col != nullptr ? vector_sum_col->info() : nullptr,
359                                                     vector_sum_row != nullptr ? vector_sum_row->info() : nullptr,
360                                                     bias != nullptr ? bias->info() : nullptr,
361                                                     output_multipliers != nullptr ? output_multipliers->info() : nullptr,
362                                                     output_shifts != nullptr ? output_shifts->info() : nullptr,
363                                                     num_elements_processed);
364     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
365     ICLKernel::configure_internal(win_config.second);
366 
367     // If _reinterpret_input_as_3d = _reinterpret_output_as_3d = true,
368     // we will dispatch a batched-GEMM to reduce the complexity of the address calculation within the OpenCL kernel.
369     // This means that the actual m used by the kernel is given by output->info()->dimension(1) and not by gemm_info.m
370     const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.m : output->info()->dimension(1);
371 
372     // Shrink M0 to be always <= M (internal_m) to prevent out-of-bounds reads.
373     // NOTE: This might have implications on heuristics and performance
374     const unsigned int internal_m0 = std::min(internal_m, lhs_info.m0);
375 
376     // Calculate partial (store instead of load) M0 and partial N0 for the partial blocks at the end of a row/column if any. This is to avoid padding.
377     const unsigned int partial_store_m0 = internal_m % internal_m0;
378     const unsigned int partial_store_n0 = gemm_info.n % rhs_info.n0;
379 
380     // Create build options
381     CLBuildOptions build_opts;
382     build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D");
383     build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
384     build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(1)));
385     build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(2)));
386     build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2)));
387     build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE");
388     build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS");
389     build_opts.add_option("-DM=" + support::cpp11::to_string(internal_m));
390     build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n));
391     build_opts.add_option("-DK=" + support::cpp11::to_string(gemm_info.k));
392     build_opts.add_option("-DM0=" + support::cpp11::to_string(internal_m0));
393     build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0));
394     build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0));
395     build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0));
396     build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0));
397     build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0));
398     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type()));
399     build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_dot8_acc_type_from_data_type(input0->info()->data_type()));
400 
401     std::string kernel_name("gemmlowp_mm_reshaped_only_rhs_");
402     kernel_name += rhs_info.transpose ? "t" : "nt";
403 
404     if(output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT)
405     {
406         kernel_name += "_fused_output_stage_fixedpoint";
407         _fuse_output_stage = true;
408         // If a_offset == 0, vector_sum_col can be a nullptr
409         if(a_offset != 0)
410         {
411             build_opts.add_option("-DA_OFFSET=" + support::cpp11::to_string(a_offset));
412             build_opts.add_option_if(vector_sum_col->info()->tensor_shape().num_dimensions() > 1, "-DSUM_COL_HAS_BATCHES");
413         }
414         // If b_offset == 0, vector_sum_row can be a nullptr
415         build_opts.add_option_if(b_offset != 0, "-DB_OFFSET=" + support::cpp11::to_string(b_offset));
416         build_opts.add_option("-DK_OFFSET=" + support::cpp11::to_string(a_offset * b_offset * input0->info()->dimension(0)));
417         build_opts.add_option_if(bias != nullptr, "-DADD_BIAS");
418         build_opts.add_option("-DRESULT_OFFSET=" + support::cpp11::to_string(output_stage.gemmlowp_offset));
419         build_opts.add_option("-DRESULT_MULTIPLIER=" + support::cpp11::to_string(output_stage.gemmlowp_multipliers[0]));
420         build_opts.add_option("-DRESULT_SHIFT=" + support::cpp11::to_string(output_stage.gemmlowp_shifts[0]));
421         build_opts.add_option_if(_is_quantized_per_channel, "-DPER_CHANNEL_QUANTIZATION");
422 
423         const int min = output_stage.gemmlowp_min_bound;
424         const int max = output_stage.gemmlowp_max_bound;
425 
426         PixelValue min_val{};
427         PixelValue max_val{};
428         std::tie(min_val, max_val) = get_min_max(output->info()->data_type());
429         build_opts.add_option_if(min != min_val.get<int32_t>(), "-DMIN_BOUND=" + support::cpp11::to_string(min));
430         build_opts.add_option_if(max != max_val.get<int32_t>(), "-DMAX_BOUND=" + support::cpp11::to_string(max));
431     }
432 
433     // Create kernel
434     _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
435 
436     // Set config_id for enabling LWS tuning
437     _config_id = kernel_name;
438     _config_id += "_";
439     _config_id += dot8_supported(CLKernelLibrary::get().get_device()) ? "_dot8" : "";
440     _config_id += "_";
441     _config_id += (_reinterpret_input_as_3d ? "3di_" : "");
442     _config_id += (_reinterpret_output_as_3d ? "3do_" : "");
443     _config_id += support::cpp11::to_string(output->info()->dimension(1));
444     _config_id += "_";
445     _config_id += support::cpp11::to_string(output->info()->dimension(0));
446     _config_id += "_";
447     _config_id += support::cpp11::to_string(gemm_info.k);
448     _config_id += "_";
449     _config_id += support::cpp11::to_string(output->info()->dimension(2));
450     _config_id += "_";
451     _config_id += support::cpp11::to_string(lhs_info.m0);
452     _config_id += "_";
453     _config_id += support::cpp11::to_string(rhs_info.n0);
454     _config_id += "_";
455     _config_id += support::cpp11::to_string(rhs_info.k0);
456     _config_id += "_";
457     _config_id += support::cpp11::to_string(rhs_info.h0);
458     _config_id += "_";
459     _config_id += support::cpp11::to_string(rhs_info.interleave);
460     ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
461 }
462 
validate(const ITensorInfo * input0,const ITensorInfo * input1,const ITensorInfo * output,const GEMMKernelInfo & gemm_info,const ITensorInfo * vector_sum_col,const ITensorInfo * vector_sum_row,const ITensorInfo * bias,const ITensorInfo * output_multipliers,const ITensorInfo * output_shifts)463 Status CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, const GEMMKernelInfo &gemm_info,
464                                                                const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias,
465                                                                const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
466 {
467     ElementsProcessed num_elements_processed{};
468     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output, gemm_info, vector_sum_col, vector_sum_row, bias, output_multipliers, output_shifts));
469     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(),
470                                                               input1->clone().get(),
471                                                               output->clone().get(),
472                                                               gemm_info,
473                                                               vector_sum_col != nullptr ? vector_sum_col->clone().get() : nullptr,
474                                                               vector_sum_row != nullptr ? vector_sum_row->clone().get() : nullptr,
475                                                               bias != nullptr ? bias->clone().get() : nullptr,
476                                                               output_multipliers != nullptr ? output_multipliers->clone().get() : nullptr,
477                                                               output_shifts != nullptr ? output_shifts->clone().get() : nullptr,
478                                                               num_elements_processed)
479                                 .first);
480 
481     return Status{};
482 }
483 
run(const Window & window,cl::CommandQueue & queue)484 void CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::run(const Window &window, cl::CommandQueue &queue)
485 {
486     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
487     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
488 
489     if(_input1->info()->num_dimensions() < 3)
490     {
491         // The stride_z for matrix B must be zero if we do not slice
492         ARM_COMPUTE_ERROR_ON(_input1->info()->strides_in_bytes()[3] != 0);
493     }
494 
495     Window slice          = window.first_slice_window_3D();
496     Window slice_matrix_b = slice;
497 
498     slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1));
499     slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1));
500 
501     if(_reinterpret_input_as_3d)
502     {
503         // Pass bottom paddings to the kernel if the input has to be reinterpreted as 3D tensor
504         const unsigned int idx0                  = 3 * num_arguments_per_2D_tensor() + 3;
505         const unsigned int total_cross_plane_pad = _input0->info()->padding().top + _input0->info()->padding().bottom;
506         _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
507     }
508 
509     if(_reinterpret_output_as_3d)
510     {
511         // Pass bottom paddings to the kernel if the output has to be reinterpreted as 3D tensor
512         const unsigned int idx0                  = 3 * num_arguments_per_2D_tensor() + 3 + (_reinterpret_input_as_3d ? 1 : 0);
513         const unsigned int total_cross_plane_pad = _output->info()->padding().top + _output->info()->padding().bottom;
514         _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
515     }
516 
517     // Set window for vector_sum_col
518     Window win_vector_sum_col = slice;
519     win_vector_sum_col.set(Window::DimY, Window::Dimension(0, 0, 0));
520     win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0));
521 
522     // Set window for vector_sum_row
523     Window win_vector_sum_row = slice;
524     win_vector_sum_row.set(Window::DimX, Window::Dimension(0, 0, 0));
525     win_vector_sum_row.set(Window::DimY, Window::Dimension(0, 0, 0));
526     win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0));
527 
528     Window biases_slice = slice;
529     biases_slice.set(Window::DimY, Window::Dimension(0, 1, 1));
530     biases_slice.set(Window::DimZ, Window::Dimension(0, 1, 1));
531 
532     do
533     {
534         Window slice_b = slice;
535         // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2
536         // This scenario can happen when the matrix multiplication is used to perform a convolution operation
537         if(!_slide_matrix_b)
538         {
539             slice_b = slice_matrix_b;
540         }
541 
542         unsigned int idx = 0;
543         add_2D_tensor_argument(idx, _input0, slice);
544         add_2D_tensor_argument(idx, _input1, slice_b);
545         add_2D_tensor_argument(idx, _output, slice);
546         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2]));
547         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2]));
548         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2]));
549         if(_reinterpret_input_as_3d)
550         {
551             // Pass bottom paddings to the kernel if the input has to be reinterpreted as 3D tensor
552             idx++;
553         }
554 
555         if(_reinterpret_output_as_3d)
556         {
557             // Pass bottom paddings to the kernel if the output has to be reinterpreted as 3D tensor
558             idx++;
559         }
560 
561         if(_fuse_output_stage)
562         {
563             add_2D_tensor_argument_if((_vector_sum_col != nullptr), idx, _vector_sum_col, win_vector_sum_col);
564             add_2D_tensor_argument_if((_vector_sum_row != nullptr), idx, _vector_sum_row, win_vector_sum_row);
565             add_1D_tensor_argument_if((_bias != nullptr), idx, _bias, biases_slice);
566             add_1D_tensor_argument_if(_is_quantized_per_channel, idx, _output_multipliers, biases_slice);
567             add_1D_tensor_argument_if(_is_quantized_per_channel, idx, _output_shifts, biases_slice);
568         }
569         enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items);
570     }
571     while(window.slide_window_slice_3D(slice));
572 }
573 } // namespace arm_compute
574