1 /*
2 * Copyright (c) 2019-2022 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/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
36 #include "src/core/AccessWindowStatic.h"
37 #include "src/core/helpers/AutoConfiguration.h"
38 #include "src/core/helpers/WindowHelpers.h"
39
40 #include "support/Cast.h"
41 #include "support/StringSupport.h"
42
43 #include <tuple>
44
45 namespace arm_compute
46 {
47 namespace opencl
48 {
49 namespace kernels
50 {
51 using namespace misc::shape_calculator;
52
53 namespace
54 {
55 using ElementsProcessed = Steps;
56
validate_arguments(const ITensorInfo * src0,const ITensorInfo * src1,const ITensorInfo * dst,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)57 Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst, const GEMMKernelInfo &gemm_info,
58 const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias,
59 const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
60 {
61 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src0, src1, dst);
62 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src0, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
63 if(src0->data_type() == DataType::QASYMM8)
64 {
65 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, src1);
66 }
67 else
68 {
69 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src1, 1, DataType::QASYMM8, DataType::QSYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8_PER_CHANNEL);
70 }
71 ARM_COMPUTE_RETURN_ERROR_ON_MSG(src0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4");
72 ARM_COMPUTE_RETURN_ERROR_ON_MSG(src1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3");
73
74 const GEMMRHSMatrixInfo rhs_info = gemm_info.rhs_info;
75 const GEMMLHSMatrixInfo lhs_info = gemm_info.lhs_info;
76 const GEMMLowpOutputStageInfo output_stage = gemm_info.output_stage;
77
78 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");
79 ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 1 || lhs_info.m0 > 8);
80 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");
81 ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs_info.export_to_cl_image, "Export to CLImage not supported for quantized GEMM");
82
83 const int m = gemm_info.m;
84 const int n = gemm_info.n;
85 const int k = gemm_info.k;
86
87 TensorShape tensor_shape1{ src1->tensor_shape() };
88 tensor_shape1.set(0, n);
89 tensor_shape1.set(1, k);
90
91 const TensorInfo tensor_info1 = src1->clone()->set_tensor_shape(tensor_shape1);
92 const TensorInfo tensor_info_reshaped1 = src1->clone()->set_tensor_shape(compute_rhs_reshaped_shape(tensor_info1, rhs_info));
93
94 ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(0) != static_cast<unsigned int>(k));
95 if(gemm_info.reinterpret_input_as_3d)
96 {
97 ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(1) * src0->dimension(2) != static_cast<unsigned int>(m));
98 }
99 else
100 {
101 ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(1) != static_cast<unsigned int>(m));
102 }
103 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src1, &tensor_info_reshaped1);
104
105 const TensorShape expected_dst_shape = compute_mm_shape(*src0, *src1, gemm_info);
106 if(dst->total_size() != 0)
107 {
108 const TensorInfo tensor_info_dst = dst->clone()->set_tensor_shape(expected_dst_shape);
109 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &tensor_info_dst);
110 if(output_stage.type == GEMMLowpOutputStageType::NONE)
111 {
112 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::S32);
113 }
114 else
115 {
116 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, dst);
117 }
118 }
119
120 if(bias != nullptr)
121 {
122 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32);
123 ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
124 ARM_COMPUTE_RETURN_ERROR_ON(expected_dst_shape[0] != bias->dimension(0));
125 }
126
127 ARM_COMPUTE_RETURN_ERROR_ON_MSG((output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN) || (output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FLOAT),
128 "Only GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT is supported");
129
130 // Checks performed if the dst stage needs to be fused
131 if(output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT)
132 {
133 // If a_offset == 0, vector_sum_col can be a nullptr
134 if(gemm_info.a_offset != 0)
135 {
136 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_col, 1, DataType::S32);
137 ARM_COMPUTE_RETURN_ERROR_ON(vector_sum_col->dimension(0) != expected_dst_shape[0]);
138 }
139
140 // If b_offset == 0, vector_sum_row can be a nullptr
141 if(gemm_info.b_offset != 0)
142 {
143 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_row, 1, DataType::S32);
144
145 // Check if mm result is a 3D reinterpretation
146 const bool reinterpret_as_3d = expected_dst_shape.num_dimensions() > 1 && expected_dst_shape.y() != vector_sum_row->tensor_shape().x();
147
148 // Validate input
149 ARM_COMPUTE_RETURN_ERROR_ON(reinterpret_as_3d && vector_sum_row->dimension(0) != (expected_dst_shape[1] * expected_dst_shape[2]));
150 ARM_COMPUTE_RETURN_ERROR_ON(!reinterpret_as_3d && vector_sum_row->dimension(0) != expected_dst_shape[1]);
151
152 if(expected_dst_shape.num_dimensions() > 1)
153 {
154 const unsigned int dst_batch_idx = reinterpret_as_3d ? 3 : 2;
155
156 TensorShape vector_sum_row_shape = vector_sum_row->tensor_shape();
157 vector_sum_row_shape.collapse_from(1);
158 TensorShape collapsed_dst_shape(expected_dst_shape);
159 collapsed_dst_shape.collapse_from(dst_batch_idx);
160
161 ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_row_shape[1] != collapsed_dst_shape[dst_batch_idx],
162 "vector_sum_row must have the same number of batches of dst tensor");
163
164 if(gemm_info.a_offset != 0)
165 {
166 TensorShape vector_sum_col_shape = vector_sum_col->tensor_shape();
167 vector_sum_col_shape.collapse_from(1);
168
169 ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_col_shape[1] != 1 && vector_sum_col_shape[1] != vector_sum_row_shape[1],
170 "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");
171 }
172 }
173 }
174
175 if(dst->total_size() != 0)
176 {
177 ARM_COMPUTE_RETURN_ERROR_ON(output_stage.output_data_type != dst->data_type());
178 }
179 ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_min_bound > output_stage.gemmlowp_max_bound);
180
181 if(output_multipliers != nullptr && output_shifts != nullptr)
182 {
183 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_multipliers, 1, DataType::S32);
184 ARM_COMPUTE_RETURN_ERROR_ON(output_multipliers->num_dimensions() > 1);
185 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_shifts, 1, DataType::S32);
186 ARM_COMPUTE_RETURN_ERROR_ON(output_shifts->num_dimensions() > 1);
187 if(output_stage.is_quantized_per_channel)
188 {
189 ARM_COMPUTE_RETURN_ERROR_ON(expected_dst_shape[0] != output_shifts->dimension(0));
190 ARM_COMPUTE_RETURN_ERROR_ON(expected_dst_shape[0] != output_multipliers->dimension(0));
191 }
192 }
193 }
194 return Status{};
195 }
196
validate_and_configure_window(const ITensorInfo * src0,const ITensorInfo * src1,ITensorInfo * dst,const GEMMKernelInfo & gemm_info,ITensorInfo * vector_sum_col,const ITensorInfo * vector_sum_row,ITensorInfo * bias,ITensorInfo * output_multipliers,ITensorInfo * output_shifts,ElementsProcessed & num_elements_processed)197 std::pair<Status, Window> validate_and_configure_window(const ITensorInfo *src0, const ITensorInfo *src1, ITensorInfo *dst, const GEMMKernelInfo &gemm_info,
198 ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, ITensorInfo *bias,
199 ITensorInfo *output_multipliers, ITensorInfo *output_shifts, ElementsProcessed &num_elements_processed)
200 {
201 const GEMMLowpOutputStageInfo output_stage = gemm_info.output_stage;
202
203 unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
204 unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
205 bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d;
206 bool reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d != 0);
207
208 Window win{};
209 Window win_out{};
210 bool window_changed = false;
211
212 // In case both input and dst have to be reinterpreted as 3D tensors,
213 // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
214 if(reinterpret_input_as_3d == reinterpret_output_as_3d)
215 {
216 reinterpret_output_as_3d = false;
217 }
218
219 // dst tensor auto initialization if not yet initialized
220 const TensorShape expected_dst_shape = compute_mm_shape(*src0, *src1, gemm_info);
221 if(output_stage.type != GEMMLowpOutputStageType::NONE)
222 {
223 auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(expected_dst_shape).set_data_type(output_stage.output_data_type));
224 }
225 else
226 {
227 auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(expected_dst_shape).set_data_type(DataType::S32));
228 }
229
230 TensorInfo tmp_info(*dst);
231
232 if(reinterpret_output_as_3d)
233 {
234 // Since the dst tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM,
235 // the window needs to be constructed on the 2D collapsed version of the tensor
236 TensorShape tmp_shape(dst->tensor_shape());
237 tmp_shape.collapse(2U, 1U);
238 tmp_info.set_tensor_shape(tmp_shape);
239 }
240
241 // Configure kernel window
242 num_elems_processed_per_iteration_x = gemm_info.rhs_info.n0;
243 num_elems_processed_per_iteration_y = gemm_info.lhs_info.m0;
244
245 win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
246 win_out = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
247
248 if(output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT)
249 {
250 if(gemm_info.a_offset != 0)
251 {
252 AccessWindowHorizontal vector_sum_col_access(vector_sum_col, 0, num_elems_processed_per_iteration_x);
253 window_changed = window_changed || update_window_and_padding(win_out, vector_sum_col_access);
254 }
255 // No access window needed for vector_sum_row
256 ARM_COMPUTE_UNUSED(vector_sum_row);
257
258 if(bias != nullptr)
259 {
260 AccessWindowHorizontal bias_access(bias, 0, num_elems_processed_per_iteration_x);
261 window_changed = window_changed || update_window_and_padding(win_out, bias_access);
262 }
263
264 if(output_multipliers != nullptr && output_stage.is_quantized_per_channel)
265 {
266 AccessWindowHorizontal output_multipliers_access(output_multipliers, 0, num_elems_processed_per_iteration_x);
267 AccessWindowHorizontal output_shifts_access(output_shifts, 0, num_elems_processed_per_iteration_x);
268 window_changed = window_changed || update_window_and_padding(win_out, output_multipliers_access, output_shifts_access);
269 }
270 }
271
272 // Collapse along the Z direction
273 // This collapse needs to be here in order to tune the Z dimension of LWS
274 Window collapsed = win;
275 const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(dst->num_dimensions()), 2u);
276 collapsed = win.collapse(win, dimension_to_collapse);
277
278 Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
279 return std::make_pair(err, collapsed);
280 }
281 } // namespace
282
ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel()283 ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel::ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel()
284 {
285 _type = CLKernelType::GEMM;
286 }
287
configure(const CLCompileContext & compile_context,const ITensorInfo * src0,const ITensorInfo * src1,ITensorInfo * dst,const GEMMKernelInfo & gemm_info,ITensorInfo * vector_sum_col,const ITensorInfo * vector_sum_row,ITensorInfo * bias,ITensorInfo * output_multipliers,ITensorInfo * output_shifts)288 void ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *src0, const ITensorInfo *src1, ITensorInfo *dst,
289 const GEMMKernelInfo &gemm_info,
290 ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, ITensorInfo *bias,
291 ITensorInfo *output_multipliers, ITensorInfo *output_shifts)
292 {
293 ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst);
294 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src0, src1, dst, gemm_info, vector_sum_col, vector_sum_row, bias, output_multipliers, output_shifts));
295
296 auto padding_info = get_padding_info({ src0, src1, dst, vector_sum_row });
297 const GEMMRHSMatrixInfo rhs_info = gemm_info.rhs_info;
298 const GEMMLHSMatrixInfo lhs_info = gemm_info.lhs_info;
299 const GEMMLowpOutputStageInfo output_stage = gemm_info.output_stage;
300 const int32_t a_offset = gemm_info.a_offset;
301 const int32_t b_offset = gemm_info.b_offset;
302
303 _reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d;
304 _reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d != 0);
305 _use_dummy_work_items = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device());
306 _is_quantized_per_channel = output_stage.is_quantized_per_channel;
307
308 // In case both input and dst have to be reinterpreted as 3D tensors,
309 // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
310 if(_reinterpret_input_as_3d == _reinterpret_output_as_3d)
311 {
312 _reinterpret_input_as_3d = false;
313 _reinterpret_output_as_3d = false;
314 }
315
316 // Check if we need to slide the matrix B
317 const unsigned int num_dimensions_src0 = src0->num_dimensions();
318 _slide_matrix_b = (src1->num_dimensions() >= num_dimensions_src0);
319
320 ElementsProcessed num_elements_processed{};
321
322 // Configure kernel window
323 auto win_config = validate_and_configure_window(src0, src1, dst, gemm_info, vector_sum_col, vector_sum_row, bias, output_multipliers, output_shifts, num_elements_processed);
324 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
325 ICLKernel::configure_internal(win_config.second);
326
327 // If _reinterpret_input_as_3d = _reinterpret_output_as_3d = true,
328 // we will dispatch a batched-GEMM to reduce the complexity of the address calculation within the OpenCL kernel.
329 // This means that the actual m used by the kernel is given by dst->dimension(1) and not by gemm_info.m
330 const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.m : dst->dimension(1);
331
332 // Shrink M0 to be always <= M (internal_m) to prevent out-of-bounds reads.
333 // NOTE: This might have implications on heuristics and performance
334 const unsigned int internal_m0 = std::min(internal_m, lhs_info.m0);
335
336 // 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.
337 const unsigned int partial_store_m0 = internal_m % internal_m0;
338 const unsigned int partial_store_n0 = gemm_info.n % rhs_info.n0;
339
340 // Create build options
341 CLBuildOptions build_opts;
342 build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D");
343 build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
344 build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(dst->dimension(1)));
345 build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(dst->dimension(2)));
346 build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(src1->dimension(2)));
347 build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE");
348 build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS");
349 build_opts.add_option("-DM=" + support::cpp11::to_string(internal_m));
350 build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n));
351 build_opts.add_option("-DK=" + support::cpp11::to_string(gemm_info.k));
352 build_opts.add_option("-DM0=" + support::cpp11::to_string(internal_m0));
353 build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0));
354 build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0));
355 build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0));
356 build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0));
357 build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0));
358 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src0->data_type()));
359 build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_dot8_acc_type_from_data_type(src0->data_type()));
360
361 std::string kernel_name("gemmlowp_mm_reshaped_only_rhs_");
362 kernel_name += rhs_info.transpose ? "t" : "nt";
363
364 if(output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT)
365 {
366 kernel_name += "_fused_output_stage_fixedpoint";
367 _fuse_output_stage = true;
368 // If a_offset == 0, vector_sum_col can be a nullptr
369 if(a_offset != 0 && vector_sum_col != nullptr)
370 {
371 build_opts.add_option("-DA_OFFSET=" + support::cpp11::to_string(a_offset));
372 build_opts.add_option_if(vector_sum_col->tensor_shape().num_dimensions() > 1, "-DSUM_COL_HAS_BATCHES");
373 }
374 // If b_offset == 0, vector_sum_row can be a nullptr
375 build_opts.add_option_if(b_offset != 0, "-DB_OFFSET=" + support::cpp11::to_string(b_offset));
376 build_opts.add_option("-DK_OFFSET=" + support::cpp11::to_string(a_offset * b_offset * src0->dimension(0)));
377 build_opts.add_option_if(bias != nullptr, "-DADD_BIAS");
378 build_opts.add_option("-DRESULT_OFFSET=" + support::cpp11::to_string(output_stage.gemmlowp_offset));
379 // In case of _is_quantized_per_channel, RESULT_MULTIPLIER and RESULT_SHIFT are not utilized, but they are passed as a part of T_QUANTIZE8 macro.
380 if(!_is_quantized_per_channel)
381 {
382 build_opts.add_option("-DRESULT_MULTIPLIER=" + support::cpp11::to_string(output_stage.gemmlowp_multipliers[0]));
383 build_opts.add_option("-DRESULT_SHIFT=" + support::cpp11::to_string(output_stage.gemmlowp_shifts[0]));
384 }
385 else
386 {
387 build_opts.add_option("-DRESULT_MULTIPLIER=0");
388 build_opts.add_option("-DRESULT_SHIFT=0");
389 }
390 build_opts.add_option_if(_is_quantized_per_channel, "-DPER_CHANNEL_QUANTIZATION");
391
392 const int min = output_stage.gemmlowp_min_bound;
393 const int max = output_stage.gemmlowp_max_bound;
394
395 PixelValue min_val{};
396 PixelValue max_val{};
397 std::tie(min_val, max_val) = get_min_max(dst->data_type());
398 build_opts.add_option_if(min != min_val.get<int32_t>(), "-DMIN_BOUND=" + support::cpp11::to_string(min));
399 build_opts.add_option_if(max != max_val.get<int32_t>(), "-DMAX_BOUND=" + support::cpp11::to_string(max));
400 }
401
402 // A macro guard to compile ONLY the kernel of interest
403 build_opts.add_option("-D" + upper_string(kernel_name));
404
405 // Create kernel
406 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
407
408 // Set config_id for enabling LWS tuning
409 _config_id = kernel_name;
410 _config_id += "_";
411 _config_id += dot8_supported(CLKernelLibrary::get().get_device()) ? "_dot8" : "";
412 _config_id += "_";
413 _config_id += (_reinterpret_input_as_3d ? "3di_" : "");
414 _config_id += (_reinterpret_output_as_3d ? "3do_" : "");
415 _config_id += support::cpp11::to_string(dst->dimension(1));
416 _config_id += "_";
417 _config_id += support::cpp11::to_string(dst->dimension(0));
418 _config_id += "_";
419 _config_id += support::cpp11::to_string(gemm_info.k);
420 _config_id += "_";
421 _config_id += support::cpp11::to_string(dst->dimension(2));
422 _config_id += "_";
423 _config_id += support::cpp11::to_string(lhs_info.m0);
424 _config_id += "_";
425 _config_id += support::cpp11::to_string(rhs_info.n0);
426 _config_id += "_";
427 _config_id += support::cpp11::to_string(rhs_info.k0);
428 _config_id += "_";
429 _config_id += support::cpp11::to_string(rhs_info.h0);
430 _config_id += "_";
431 _config_id += support::cpp11::to_string(rhs_info.interleave);
432 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
433 }
434
validate(const ITensorInfo * src0,const ITensorInfo * src1,const ITensorInfo * dst,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)435 Status ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel::validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst, const GEMMKernelInfo &gemm_info,
436 const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias,
437 const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
438 {
439 ElementsProcessed num_elements_processed{};
440 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src0, src1, dst, gemm_info, vector_sum_col, vector_sum_row, bias, output_multipliers, output_shifts));
441 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src0->clone().get(),
442 src1->clone().get(),
443 dst->clone().get(),
444 gemm_info,
445 vector_sum_col != nullptr ? vector_sum_col->clone().get() : nullptr,
446 vector_sum_row != nullptr ? vector_sum_row->clone().get() : nullptr,
447 bias != nullptr ? bias->clone().get() : nullptr,
448 output_multipliers != nullptr ? output_multipliers->clone().get() : nullptr,
449 output_shifts != nullptr ? output_shifts->clone().get() : nullptr,
450 num_elements_processed)
451 .first);
452
453 return Status{};
454 }
455
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)456 void ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
457 {
458 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
459 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
460
461 const auto src0 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
462 const auto src1 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
463 const auto bias = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_BIAS));
464 const auto vector_sum_col = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_VEC_COL_SUM));
465 const auto vector_sum_row = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_VEC_ROW_SUM));
466 const auto output_shifts = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SHIFTS));
467 const auto output_multipliers = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_MULTIPLIERS));
468 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
469
470 if(src1->info()->num_dimensions() < 3)
471 {
472 // The stride_z for matrix B must be zero if we do not slice
473 ARM_COMPUTE_ERROR_ON(src1->info()->strides_in_bytes()[3] != 0);
474 }
475
476 Window slice = window.first_slice_window_3D();
477 Window slice_matrix_b = slice;
478
479 slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1));
480 slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1));
481
482 if(_reinterpret_input_as_3d)
483 {
484 // Pass bottom paddings to the kernel if the input has to be reinterpreted as 3D tensor
485 const unsigned int idx0 = 3 * num_arguments_per_2D_tensor() + 3;
486 const unsigned int total_cross_plane_pad = src0->info()->padding().top + src0->info()->padding().bottom;
487 _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
488 }
489
490 if(_reinterpret_output_as_3d)
491 {
492 // Pass bottom paddings to the kernel if the dst has to be reinterpreted as 3D tensor
493 const unsigned int idx0 = 3 * num_arguments_per_2D_tensor() + 3 + (_reinterpret_input_as_3d ? 1 : 0);
494 const unsigned int total_cross_plane_pad = dst->info()->padding().top + dst->info()->padding().bottom;
495 _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
496 }
497
498 // Set window for vector_sum_col
499 Window win_vector_sum_col = slice;
500 win_vector_sum_col.set(Window::DimY, Window::Dimension(0, 0, 0));
501 win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0));
502
503 // Set window for vector_sum_row
504 Window win_vector_sum_row = slice;
505 win_vector_sum_row.set(Window::DimX, Window::Dimension(0, 0, 0));
506 win_vector_sum_row.set(Window::DimY, Window::Dimension(0, 0, 0));
507 win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0));
508
509 Window biases_slice = slice;
510 biases_slice.set(Window::DimY, Window::Dimension(0, 1, 1));
511 biases_slice.set(Window::DimZ, Window::Dimension(0, 1, 1));
512
513 do
514 {
515 Window slice_b = slice;
516 // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2
517 // This scenario can happen when the matrix multiplication is used to perform a convolution operation
518 if(!_slide_matrix_b)
519 {
520 slice_b = slice_matrix_b;
521 }
522
523 unsigned int idx = 0;
524 add_2D_tensor_argument(idx, src0, slice);
525 add_2D_tensor_argument(idx, src1, slice_b);
526 add_2D_tensor_argument(idx, dst, slice);
527 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src0->info()->strides_in_bytes()[2]));
528 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src1->info()->strides_in_bytes()[2]));
529 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(dst->info()->strides_in_bytes()[2]));
530 if(_reinterpret_input_as_3d)
531 {
532 // Pass bottom paddings to the kernel if the input has to be reinterpreted as 3D tensor
533 idx++;
534 }
535
536 if(_reinterpret_output_as_3d)
537 {
538 // Pass bottom paddings to the kernel if the dst has to be reinterpreted as 3D tensor
539 idx++;
540 }
541
542 if(_fuse_output_stage)
543 {
544 add_2D_tensor_argument_if((vector_sum_col != nullptr), idx, vector_sum_col, win_vector_sum_col);
545 add_2D_tensor_argument_if((vector_sum_row != nullptr), idx, vector_sum_row, win_vector_sum_row);
546 add_1D_tensor_argument_if((bias != nullptr), idx, bias, biases_slice);
547 add_1D_tensor_argument_if(_is_quantized_per_channel, idx, output_multipliers, biases_slice);
548 add_1D_tensor_argument_if(_is_quantized_per_channel, idx, output_shifts, biases_slice);
549 }
550 enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items);
551 }
552 while(window.slide_window_slice_3D(slice));
553 }
554 } // namespace kernels
555 } // namespace opencl
556 } // namespace arm_compute
557