aboutsummaryrefslogtreecommitdiff
path: root/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.cpp
blob: 9630caefd88b5b363da875e93e047d7b8000c5b8 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
/*
 * Copyright (c) 2018-2019 ARM Limited.
 *
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a copy
 * of this software and associated documentation files (the "Software"), to
 * deal in the Software without restriction, including without limitation the
 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
 * sell copies of the Software, and to permit persons to whom the Software is
 * furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in all
 * copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 */
#include "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h"

#include "arm_compute/core/AccessWindowStatic.h"
#include "arm_compute/core/CL/CLHelpers.h"
#include "arm_compute/core/CL/CLKernelLibrary.h"
#include "arm_compute/core/CL/CLValidate.h"
#include "arm_compute/core/CL/ICLTensor.h"
#include "arm_compute/core/CL/OpenCL.h"
#include "arm_compute/core/Error.h"
#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/TensorInfo.h"
#include "arm_compute/core/Types.h"
#include "arm_compute/core/Utils.h"
#include "arm_compute/core/Validate.h"
#include "arm_compute/core/Window.h"
#include "arm_compute/core/utils/helpers/float_ops.h"
#include "arm_compute/core/utils/misc/ShapeCalculator.h"
#include "support/ToolchainSupport.h"

#include <cstddef>
#include <cstdint>
#include <tuple>

using namespace arm_compute;
using namespace arm_compute::misc::shape_calculator;

namespace arm_compute
{
class Coordinates;
} // namespace arm_compute

namespace
{
using ElementsProcessed = Steps;

Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info,
                          const GEMMRHSMatrixInfo &rhs_info,
                          const GEMMKernelInfo    &gemm_info)
{
    ARM_COMPUTE_UNUSED(alpha);
    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output);
    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input0);
    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F32);
    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4");
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3");
    ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 != rhs_info.k0);
    ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.transpose == rhs_info.transpose);
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(((lhs_info.k0 & (lhs_info.k0 - 1)) && lhs_info.k0 != 3), "Only 2,3,4,8,16 are supported for k0");
    ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 > 16);
    ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 2 || lhs_info.m0 > 8);
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.n0 & (rhs_info.n0 - 1)) && rhs_info.n0 != 3), "Only 2,3,4,8,16 are supported for n0");
    ARM_COMPUTE_RETURN_ERROR_ON_MSG((gemm_info.reinterpret_input_as_3d || gemm_info.depth_output_gemm3d != 0) && (input2 != nullptr)
                                    && (!gemm_info.broadcast_bias),
                                    "Bias addition only supported with broadcast mode in case the input or output has to be reinterpreted as 3D");

    const unsigned int m = gemm_info.m;
    const unsigned int n = gemm_info.n;
    const unsigned int k = gemm_info.k;

    TensorShape tensor_shape0{ input0->tensor_shape() };
    tensor_shape0.set(0, k);
    tensor_shape0.set(1, m);

    TensorShape tensor_shape1{ input1->tensor_shape() };
    tensor_shape1.set(0, n);
    tensor_shape1.set(1, k);

    if(input2 != nullptr && !(helpers::float_ops::is_zero(beta)))
    {
        const unsigned int input2_dim0 = input2->dimension(0);
        const unsigned int input2_dim1 = input2->dimension(1);

        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input2, input1);
        if(gemm_info.broadcast_bias)
        {
            ARM_COMPUTE_RETURN_ERROR_ON_MSG((input2_dim1 != 1 || input2_dim0 != n), "Incorrect dimension of bias matrix which is to be broadcasted");
        }
        else
        {
            ARM_COMPUTE_RETURN_ERROR_ON_MSG((input2_dim0 != n || input2_dim1 != m), "Incorrect dimension of bias matrix");
        }
    }

    const TensorInfo tensor_info0 = input0->clone()->set_tensor_shape(tensor_shape0);
    const TensorInfo tensor_info1 = input1->clone()->set_tensor_shape(tensor_shape1);

    const TensorInfo tensor_info_reshaped0 = input0->clone()->set_tensor_shape(compute_lhs_reshaped_shape(tensor_info0, lhs_info));
    const TensorInfo tensor_info_reshaped1 = input1->clone()->set_tensor_shape(compute_rhs_reshaped_shape(tensor_info1, rhs_info));

    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input0, &tensor_info_reshaped0);
    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, &tensor_info_reshaped1);

    if(output->total_size() != 0)
    {
        const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info));
        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output);
    }

    return Status{};
}

std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info,
                                                        const GEMMRHSMatrixInfo &rhs_info,
                                                        const GEMMKernelInfo &gemm_info, ElementsProcessed &num_elements_processed)
{
    unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
    unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
    bool          reinterpret_output_as_3d            = gemm_info.depth_output_gemm3d != 0;

    Window win{};
    Window win_out{};
    bool   window_changed = false;

    // Output tensor auto initialization if not yet initialized
    auto_init_if_empty(*output, input0->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info)));

    TensorInfo tmp_info(*output);

    if(reinterpret_output_as_3d)
    {
        // Since the output tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM,
        // the window needs to be constructed on the 2D collapsed version of the tensor
        TensorShape tmp_shape(output->tensor_shape());
        tmp_shape.collapse(2U, 1U);
        tmp_info.set_tensor_shape(tmp_shape);
    }

    // Configure kernel window
    num_elems_processed_per_iteration_x = rhs_info.n0;
    num_elems_processed_per_iteration_y = lhs_info.m0;

    // Note: bottom paddings are calculated manually as the output can be reinterpreted as 3D tensor
    // The only way to set properly the paddings, it is to set those explicitly through the AccessWindowStatic
    const unsigned int m          = gemm_info.m;
    const unsigned int bottom_pad = (num_elems_processed_per_iteration_y - (m % num_elems_processed_per_iteration_y)) % num_elems_processed_per_iteration_y;

    win     = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
    win_out = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));

    AccessWindowStatic input0_access(input0, 0, 0,
                                     ceil_to_multiple(input0->dimension(0), num_elems_processed_per_iteration_y),
                                     input0->dimension(1));
    AccessWindowStatic input1_access(input1, 0, 0,
                                     ceil_to_multiple(input1->dimension(0), num_elems_processed_per_iteration_x),
                                     input1->dimension(1));
    AccessWindowStatic output_access(output, 0, 0,
                                     ceil_to_multiple(output->dimension(0), num_elems_processed_per_iteration_x),
                                     output->dimension(1) + bottom_pad);

    if(input2 != nullptr)
    {
        const int bias_processed_per_iteration_x = num_elems_processed_per_iteration_x;

        const int bias_processed_per_iteration_y = gemm_info.broadcast_bias ? 1 : num_elems_processed_per_iteration_y;

        AccessWindowStatic input2_access(input2, 0, 0,
                                         ceil_to_multiple(input2->dimension(0), bias_processed_per_iteration_x),
                                         ceil_to_multiple(input2->dimension(1), bias_processed_per_iteration_y));

        window_changed = update_window_and_padding(win, input0_access, input1_access, input2_access) || // window used by the execute_window_loop
                         update_window_and_padding(win_out, output_access);                             // window used to update the padding requirements of output tensor
    }
    else
    {
        window_changed = update_window_and_padding(win, input0_access, input1_access) || // window used by the execute_window_loop
                         update_window_and_padding(win_out, output_access);              // window used to update the padding requirements of output tensor
    }

    output_access.set_valid_region(win_out, ValidRegion(Coordinates(0, 0), output->tensor_shape()));

    // Collapse along the Z direction
    // This collapse needs to be here in order to tune the Z dimension of LWS
    Window             collapsed             = win;
    const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(output->num_dimensions()), 2u);
    collapsed                                = win.collapse(win, dimension_to_collapse);

    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
    return std::make_pair(err, collapsed);
}
} // namespace

CLGEMMMatrixMultiplyReshapedKernel::CLGEMMMatrixMultiplyReshapedKernel()
    : _input0(nullptr), _input1(nullptr), _input2(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_output_as_3d(false), _k(1), _use_dummy_work_items(false), _add_bias(false),
      _broadcast_bias(false)
{
}

void CLGEMMMatrixMultiplyReshapedKernel::configure(const ICLTensor *input0, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float alpha, float beta,
                                                   const GEMMLHSMatrixInfo &lhs_info,
                                                   const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
{
    ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output);

    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), (input2 != nullptr ? input2->info() : nullptr), output->info(), alpha, beta, lhs_info, rhs_info, gemm_info));

    _input0                   = input0;
    _input1                   = input1;
    _input2                   = helpers::float_ops::is_zero(beta) ? nullptr : input2;
    _output                   = output;
    _reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0;
    _k                        = gemm_info.k;
    _use_dummy_work_items     = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device());
    _add_bias                 = _input2 != nullptr;
    _broadcast_bias           = gemm_info.broadcast_bias;

    // Check if we need to slide the matrix B
    const unsigned int num_dimensions_input0 = _input0->info()->num_dimensions();
    _slide_matrix_b                          = (_input1->info()->num_dimensions() >= num_dimensions_input0);

    ElementsProcessed num_elements_processed{};

    // Configure kernel window
    auto win_config = validate_and_configure_window(input0->info(), input1->info(), input2 != nullptr ? input2->info() : nullptr, output->info(), lhs_info, rhs_info, gemm_info, num_elements_processed);
    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
    ICLKernel::configure_internal(win_config.second);

    // Create build options
    CLBuildOptions build_opts;
    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type()));
    build_opts.add_option_if(!(helpers::float_ops::is_one(alpha)), "-DALPHA=" + float_to_string_with_full_precision(alpha));
    build_opts.add_option_if(_input2 != nullptr, "-DBETA=" + float_to_string_with_full_precision(beta));
    build_opts.add_option_if(helpers::float_ops::is_one(beta), "-DUNIT_BETA");
    build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
    build_opts.add_option_if(_reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(1)));
    build_opts.add_option_if(_reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(2)));
    build_opts.add_option_if(gemm_info.broadcast_bias, "-DBROADCAST_BIAS");
    build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2)));
    build_opts.add_option_if(lhs_info.interleave, "-DLHS_INTERLEAVE");
    build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE");
    build_opts.add_option_if(lhs_info.transpose, "-DLHS_TRANSPOSE");
    build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS");
    build_opts.add_option("-DM=" + support::cpp11::to_string(gemm_info.m));
    build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n));
    build_opts.add_option("-DM0=" + support::cpp11::to_string(lhs_info.m0));
    build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0));
    build_opts.add_option("-DK0=" + support::cpp11::to_string(lhs_info.k0));
    build_opts.add_option("-DV0=" + support::cpp11::to_string(lhs_info.v0));
    build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0));
    build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation())));
    build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a()));
    build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b()));

    std::string kernel_name("gemm_mm_reshaped_");
    kernel_name += lhs_info.transpose ? "lhs_t_" : "lhs_nt_";
    kernel_name += rhs_info.transpose ? "rhs_t" : "rhs_nt";

    // Create kernel
    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));

    // Set config_id for enabling LWS tuning
    _config_id = kernel_name;
    _config_id += "_";
    _config_id += (_add_bias ? "add_bias_" : "");
    _config_id += (_broadcast_bias ? "broadcast_bias_" : "");
    _config_id += (_reinterpret_output_as_3d ? "3do_" : "");
    _config_id += (gemm_info.activation_info.enabled() ? "fused_activation_" : "");
    _config_id += lower_string(string_from_data_type(input0->info()->data_type()));
    _config_id += "_";
    _config_id += support::cpp11::to_string(output->info()->dimension(1));
    _config_id += "_";
    _config_id += support::cpp11::to_string(output->info()->dimension(0));
    _config_id += "_";
    _config_id += support::cpp11::to_string(gemm_info.k);
    _config_id += "_";
    _config_id += support::cpp11::to_string(output->info()->dimension(2));
    _config_id += "_";
    _config_id += support::cpp11::to_string(lhs_info.m0);
    _config_id += "_";
    _config_id += support::cpp11::to_string(rhs_info.n0);
    _config_id += "_";
    _config_id += support::cpp11::to_string(lhs_info.k0);
    _config_id += "_";
    _config_id += support::cpp11::to_string(lhs_info.v0);
    _config_id += "_";
    _config_id += support::cpp11::to_string(rhs_info.h0);
    _config_id += "_";
    _config_id += support::cpp11::to_string(lhs_info.interleave);
    _config_id += "_";
    _config_id += support::cpp11::to_string(rhs_info.interleave);
}

Status CLGEMMMatrixMultiplyReshapedKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float alpha, float beta,
                                                    const GEMMLHSMatrixInfo &lhs_info,
                                                    const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
{
    ElementsProcessed num_elements_processed{};
    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, input2, output, alpha, beta, lhs_info, rhs_info, gemm_info));
    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(),
                                                              input1->clone().get(),
                                                              input2 != nullptr ? input2->clone().get() : nullptr,
                                                              output->clone().get(),
                                                              lhs_info,
                                                              rhs_info,
                                                              gemm_info,
                                                              num_elements_processed)
                                .first);

    return Status{};
}

void CLGEMMMatrixMultiplyReshapedKernel::run(const Window &window, cl::CommandQueue &queue)
{
    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);

    if(_input1->info()->num_dimensions() < 3)
    {
        // The stride_z for matrix B must be zero if we do not slice
        ARM_COMPUTE_ERROR_ON(_input1->info()->strides_in_bytes()[3] != 0);
    }

    Window slice          = window.first_slice_window_3D();
    Window slice_matrix_b = slice;

    slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1));
    slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1));

    if(_reinterpret_output_as_3d)
    {
        // Pass bottom paddings to the kernel if the output has to be reinterpreted as 3D tensor
        unsigned int idx0;
        if(_add_bias)
        {
            idx0 = 4 * num_arguments_per_2D_tensor() + 5;
        }
        else
        {
            idx0 = 3 * num_arguments_per_2D_tensor() + 4;
        }
        const unsigned int total_cross_plane_pad = _output->info()->padding().top + _output->info()->padding().bottom;
        _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
    }

    do
    {
        Window slice_b = slice;
        // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2
        // This scenario can happen when the matrix multiplication is used to perform a convolution operation
        if(!_slide_matrix_b)
        {
            slice_b = slice_matrix_b;
        }

        unsigned int idx = 0;
        add_2D_tensor_argument(idx, _input0, slice);
        add_2D_tensor_argument(idx, _input1, slice_b);
        add_2D_tensor_argument_if((_add_bias), idx, _input2, slice);
        add_2D_tensor_argument(idx, _output, slice);
        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_k));
        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2]));
        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2]));
        if(_add_bias)
        {
            _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input2->info()->strides_in_bytes()[2]));
        }
        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2]));
        enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items);
    }
    while(window.slide_window_slice_3D(slice));
}