aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/NEGEMMInterleave4x4Kernel.cpp
blob: a9c04824aeb1ffd581667d95f903d334da8f8fa9 (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
/*
 * Copyright (c) 2016-2019 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/NEON/kernels/NEGEMMInterleave4x4Kernel.h"

#include "arm_compute/core/Error.h"
#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/ITensor.h"
#include "arm_compute/core/NEON/INEKernel.h"
#include "arm_compute/core/Types.h"
#include "arm_compute/core/Validate.h"
#include "arm_compute/core/Window.h"
#include "arm_compute/core/utils/misc/ShapeCalculator.h"

#include <arm_neon.h>
#include <cstddef>
#include <cstdint>
#include <tuple>

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

namespace
{
Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output)
{
    //Note: ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input) is not needed here as this kernel doesn't use NEON FP16 instructions.
    ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);

    if(output->total_size() != 0)
    {
        TensorShape output_shape = input->tensor_shape();
        output_shape.set(0, input->dimension(0) * 4);
        output_shape.set(1, std::ceil(input->dimension(1) / 4.0f));
        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
    }

    return Status{};
}

std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
{
    unsigned int           num_elems_processed_per_iteration_x = (input->element_size() == 1) ? 8 : 4;
    constexpr unsigned int num_elems_processed_per_iteration_y = 4;
    bool                   window_changed                      = false;

    // Configure kernel window
    Window                win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
    AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
    window_changed = window_changed || update_window_and_padding(win, input_access);

    // Configure window in case of configured output
    if(output->total_size() != 0)
    {
        AccessWindowRectangle output_access(output, 0, 0, num_elems_processed_per_iteration_x * num_elems_processed_per_iteration_y, 1, 4.0f, 0.25f);
        window_changed = window_changed || update_window_and_padding(win, output_access);
        output_access.set_valid_region(win, input->valid_region());
    }

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

void gemm_interleave_8bit_elements(const ITensor *input, ITensor *output, const Window &window)
{
    const size_t in_stride = input->info()->strides_in_bytes()[1];

    // Set window for output tensor
    Window win_out(window);
    win_out.scale(Window::DimY, 0.25f);
    Iterator in(input, window);

    win_out.set_dimension_step(Window::DimX, 32);
    Iterator out(output, win_out);

    execute_window_loop(window, [&](const Coordinates &)
    {
        const uint8x8x4_t data =
        {
            {
                vld1_u8(in.ptr() + 0 * in_stride),
                vld1_u8(in.ptr() + 1 * in_stride),
                vld1_u8(in.ptr() + 2 * in_stride),
                vld1_u8(in.ptr() + 3 * in_stride),
            }
        };
        vst4_u8(out.ptr(), data);
    },
    in, out);
}

void gemm_interleave_16bit_elements(const ITensor *input, ITensor *output, const Window &window)
{
    const size_t in_stride = input->info()->strides_in_bytes()[1];

    // Set window for output tensor
    Window win_out(window);
    win_out.scale(Window::DimY, 0.25f);
    Iterator in(input, window);

    win_out.set_dimension_step(Window::DimX, 16);
    Iterator out(output, win_out);

    execute_window_loop(window, [&](const Coordinates &)
    {
        const uint16x4x4_t data =
        {
            {
                vld1_u16(reinterpret_cast<const uint16_t *>(in.ptr() + 0 * in_stride)),
                vld1_u16(reinterpret_cast<const uint16_t *>(in.ptr() + 1 * in_stride)),
                vld1_u16(reinterpret_cast<const uint16_t *>(in.ptr() + 2 * in_stride)),
                vld1_u16(reinterpret_cast<const uint16_t *>(in.ptr() + 3 * in_stride)),
            }
        };
        vst4_u16(reinterpret_cast<uint16_t *>(out.ptr()), data);
    },
    in, out);
}

void gemm_interleave_32bit_elements(const ITensor *input, ITensor *output, const Window &window)
{
    const size_t in_stride = input->info()->strides_in_bytes()[1];

    // Set window for output tensor
    Window win_out(window);
    win_out.scale(Window::DimY, 0.25f);
    Iterator in(input, window);

    win_out.set_dimension_step(Window::DimX, 16);
    Iterator out(output, win_out);

    execute_window_loop(window, [&](const Coordinates &)
    {
        const uint32x4x4_t data =
        {
            {
                vld1q_u32(reinterpret_cast<const uint32_t *>(in.ptr() + 0 * in_stride)),
                vld1q_u32(reinterpret_cast<const uint32_t *>(in.ptr() + 1 * in_stride)),
                vld1q_u32(reinterpret_cast<const uint32_t *>(in.ptr() + 2 * in_stride)),
                vld1q_u32(reinterpret_cast<const uint32_t *>(in.ptr() + 3 * in_stride))
            }
        };
        vst4q_u32(reinterpret_cast<uint32_t *>(out.ptr()), data);
    },
    in, out);
}
} // namespace

NEGEMMInterleave4x4Kernel::NEGEMMInterleave4x4Kernel()
    : _func(nullptr)
{
}

void NEGEMMInterleave4x4Kernel::configure(const ITensor *input, ITensor *output)
{
    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);

    // Output auto inizialitation if not yet initialized
    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(compute_interleaved_shape(*input->info())));

    // Perform validate step
    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info()));

    _input  = input;
    _output = output;

    switch(input->info()->element_size())
    {
        case 1:
            _func = &gemm_interleave_8bit_elements;
            break;
        case 2:
            _func = &gemm_interleave_16bit_elements;
            break;
        case 4:
            _func = &gemm_interleave_32bit_elements;
            break;
        default:
            ARM_COMPUTE_ERROR_ON("Element size not supported");
            break;
    }

    // Configure kernel window
    auto win_config = validate_and_configure_window(input->info(), output->info());
    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
    INEKernel::configure(win_config.second);
}

Status NEGEMMInterleave4x4Kernel::validate(const ITensorInfo *input, const ITensorInfo *output)
{
    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output));
    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);

    return Status{};
}

void NEGEMMInterleave4x4Kernel::run(const Window &window, const ThreadInfo &info)
{
    ARM_COMPUTE_UNUSED(info);
    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
    ARM_COMPUTE_ERROR_ON(_func == nullptr);
    /*
    *  This kernel puts the values in a 4x4 block of Matrix A on the same row (Interleaved values)
    *         |a00 a01 a02 a03|
    *         |a10 a11 a12 a13|
    *         |a20 a21 a22 a23| = | a00 a10 a20 a30 || a01 a11 a21 a31 || a02 a12 a22 a32 || a03 a13 a23 a33 |
    *         |a30 a31 a32 a33|
    *
    *         After this operation, the output matrix will have the following shape: [ height * 4, ceil(width / 4.0f) ]
    */
    (*_func)(_input, _output, window);
}