1 /*
2 * Copyright (c) 2020-2021 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 "arm_compute/core/Types.h"
25 #include "arm_compute/core/WindowIterator.h"
26 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
27 #include "arm_compute/runtime/NEON/NEFunctions.h"
28 #include "arm_compute/runtime/NEON/NEScheduler.h"
29 #include "support/ToolchainSupport.h"
30 #include "utils/Utils.h"
31
32 #include <cstdlib>
33
34 using namespace arm_compute;
35 using namespace utils;
36
37 // Find min and max value in a float array
find_min_max(int size,const float * data,float * min,float * max)38 void find_min_max(int size, const float *data, float *min, float *max)
39 {
40 *min = *max = data[0];
41 for(int i = 0; i < size; i++)
42 {
43 const float val = data[i];
44 *min = std::min(*min, val);
45 *max = std::max(*max, val);
46 }
47 }
48
49 // Return reasonable quantisation parameters to use for an array of floats
50 // based on min and max values
choose_quantization_params(float min,float max)51 QuantizationInfo choose_quantization_params(float min, float max)
52 {
53 // Extend the [min,max] interval to contain 0 so we can represent it exactly
54 min = std::min(min, 0.f);
55 max = std::max(max, 0.f);
56
57 // Set the quantized min and max in float values
58 const float qmin = 0;
59 const float qmax = 255;
60
61 // Determine the scale
62 const float scale = (max - min) / (qmax - qmin);
63
64 // Determine the zero-point; using affine equation val = (qval-zerop) * scale
65 const float zero_point_real = qmin - min / scale;
66
67 // But we need to nudge the zero_point to an integer (exact quantized value)
68 std::uint8_t zero_point_nudged = 0;
69 if(zero_point_real < qmin)
70 {
71 zero_point_nudged = qmin;
72 }
73 else if(zero_point_real > qmax)
74 {
75 zero_point_nudged = qmax;
76 }
77 else
78 {
79 zero_point_nudged = static_cast<std::uint8_t>(support::cpp11::round(zero_point_real));
80 }
81
82 QuantizationInfo qinfo = QuantizationInfo(scale, zero_point_nudged);
83 return qinfo;
84 }
85
quantize_values(int size,qasymm8_t * output,float * input,const QuantizationInfo qinfo)86 void quantize_values(int size, qasymm8_t *output, float *input, const QuantizationInfo qinfo)
87 {
88 for(int i = 0; i < size; i++)
89 {
90 output[i] = quantize_qasymm8(input[i], qinfo);
91 }
92 std::cout << "\n";
93 }
94
main(int argc,char ** argv)95 int main(int argc, char **argv)
96 {
97 Tensor src1;
98 Tensor src2;
99 Tensor dst0;
100 Tensor q_src1;
101 Tensor q_src2;
102 Tensor q_dst0;
103 Tensor q_res;
104 Tensor q_res_output;
105 size_t M = 4;
106 size_t N = 4;
107 size_t K = 4;
108 bool default_input = true;
109
110 // Parse args
111 if(argc < 3) /* case default matrix sizes */
112 {
113 // Print help
114 std::cout << "Usage: ./build/neon_gemm_qasymm8 M N K\n";
115 std::cout << "Too few or no inputs provided. Using default M=4, N=4, K=4\n\n";
116 }
117 else /* case M N K arguments provided */
118 {
119 M = strtol(argv[1], nullptr, 10);
120 N = strtol(argv[2], nullptr, 10);
121 K = strtol(argv[3], nullptr, 10);
122 default_input = false;
123 }
124
125 /*** Floating point matrix multiplication ***/
126
127 // Initialise input matrices
128 NEGEMM fgemm{};
129
130 src1.allocator()->init(TensorInfo(TensorShape(K, M), 1, DataType::F32));
131 src2.allocator()->init(TensorInfo(TensorShape(N, K), 1, DataType::F32));
132 dst0.allocator()->init(TensorInfo(TensorShape(N, M), 1, DataType::F32));
133 fgemm.configure(&src1, &src2, nullptr, &dst0, 1, 0);
134
135 // Allocate matrices
136 src1.allocator()->allocate();
137 src2.allocator()->allocate();
138 dst0.allocator()->allocate();
139
140 // Fill in tensors, by default fill in with known data - for easy testing
141 auto *src1_ptr = reinterpret_cast<float *>(src1.buffer());
142 auto *src2_ptr = reinterpret_cast<float *>(src2.buffer());
143 auto *dst0_ptr = reinterpret_cast<float *>(dst0.buffer());
144
145 // Fill in: one is the identity matrix, other is sequential values
146 // src1: Identity matrix
147 for(size_t i = 0; i < M * K; i++)
148 {
149 src1_ptr[i] = 0;
150 }
151 for(size_t i = 0; i < M; i++)
152 {
153 src1_ptr[i * K + i] = 1.0f;
154 }
155
156 // src2: Sequential values matrix
157 for(size_t i = 0; i < K * N; i++)
158 {
159 src2_ptr[i] = i * 1.123f;
160 }
161
162 // Otherwise if M, N, K is given, fill in with random values
163 if(!default_input)
164 {
165 fill_random_tensor(src1, 0.f, 1.f);
166 fill_random_tensor(src2, 0.f, 1.f);
167 }
168
169 // Run single precision gemm and print result
170 fgemm.run();
171
172 #if ARM_COMPUTE_DEBUG_ENABLED
173 std::cout << "Result matrix:\n";
174 src1.print(std::cout);
175 src2.print(std::cout);
176 dst0.print(std::cout);
177 #endif // ARM_COMPUTE_DEBUG_ENABLED
178
179 /*** Quantised asymmetric 8bit matrix multiplication ***/
180
181 // Start by finding the quantisation parameters for each set of values
182 float src1_min;
183 float src1_max;
184 float src2_min;
185 float src2_max;
186 float dst0_min;
187 float dst0_max;
188
189 find_min_max(M * K, src1_ptr, &src1_min, &src1_max);
190 find_min_max(K * N, src2_ptr, &src2_min, &src2_max);
191 find_min_max(M * N, dst0_ptr, &dst0_min, &dst0_max);
192
193 const QuantizationInfo src1_qinfo = choose_quantization_params(src1_min, src1_max);
194 const QuantizationInfo src2_qinfo = choose_quantization_params(src2_min, src2_max);
195 const QuantizationInfo dst0_qinfo = choose_quantization_params(dst0_min, dst0_max);
196
197 std::cout << "Matrix 1: min=" << src1_min << ", max=" << src1_max << ", ";
198 std::cout << "QuantisationInfo(" << src1_qinfo.scale()[0] << ", " << src1_qinfo.offset()[0] << ")\n";
199 std::cout << "Matrix 2: min=" << src2_min << ", max=" << src2_max << ", ";
200 std::cout << "QuantisationInfo(" << src2_qinfo.scale()[0] << ", " << src2_qinfo.offset()[0] << ")\n";
201 std::cout << "Result : min=" << dst0_min << ", max=" << dst0_max << ", ";
202 std::cout << "QuantisationInfo(" << dst0_qinfo.scale()[0] << ", " << dst0_qinfo.offset()[0] << ")\n";
203
204 // We now have the quantisation info and can configure the quantised tensors
205 q_src1.allocator()->init(TensorInfo(TensorShape(K, M), 1, DataType::QASYMM8, src1_qinfo));
206 q_src2.allocator()->init(TensorInfo(TensorShape(N, K), 1, DataType::QASYMM8, src2_qinfo));
207 q_dst0.allocator()->init(TensorInfo(TensorShape(N, M), 1, DataType::QASYMM8, dst0_qinfo));
208
209 // In this approach we use the QuantizationLayer construct to perform quantization
210 NEQuantizationLayer q1;
211 NEQuantizationLayer q2;
212 NEQuantizationLayer q3;
213 q1.configure(&src1, &q_src1);
214 q2.configure(&src2, &q_src2);
215 q3.configure(&dst0, &q_dst0);
216
217 // Configure low precision gemm and initialise result tensor (pre-output)
218 NEGEMMLowpMatrixMultiplyCore qgemm;
219 q_res.allocator()->init(TensorInfo(TensorShape(N, M), 1, DataType::S32));
220 qgemm.configure(&q_src1, &q_src2, nullptr, &q_res);
221
222 // Configure output stage after computing shift and multiplier parameters
223 NEGEMMLowpOutputStage gemmlowp_output_stage;
224 int output_multiplier;
225 int output_shift;
226 float multiplier = (src1_qinfo.uniform().scale * src2_qinfo.uniform().scale) / dst0_qinfo.uniform().scale;
227 quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
228 std::cout << "(q_multiplier, q_shift) = (" << output_multiplier << ", " << output_shift << ")\n\n";
229
230 GEMMLowpOutputStageInfo info;
231 info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
232 info.gemmlowp_multiplier = output_multiplier;
233 info.gemmlowp_shift = output_shift;
234 info.gemmlowp_offset = dst0_qinfo.uniform().offset;
235 info.output_data_type = DataType::QASYMM8;
236 q_res_output.info()->set_data_type(DataType::QASYMM8);
237 q_res_output.info()->set_num_channels(1);
238 gemmlowp_output_stage.configure(&q_res, nullptr, &q_res_output, info);
239
240 // Allocate all tensors
241 q_src1.allocator()->allocate();
242 q_src2.allocator()->allocate();
243 q_dst0.allocator()->allocate();
244 q_res.allocator()->allocate();
245 q_res_output.allocator()->allocate();
246
247 // Run quantization layers (quantizes values of each tensor)
248 q1.run();
249 q2.run();
250 q3.run();
251 // Run low precision matrix multiply kernel
252 qgemm.run();
253 // Run output stage kernel
254 gemmlowp_output_stage.run();
255 std::cout << "\nTest Passed\n";
256
257 #if ARM_COMPUTE_DEBUG_ENABLED
258 // Print quantized source matrices
259 q_src1.print(std::cout);
260 q_src2.print(std::cout);
261 // Print result matrix in int32 form - before output stage processing
262 std::cout << "Lowp GEMM output (int32):\n";
263 q_res.print(std::cout);
264 // Print QASYMM8 (quantized) matrix
265 std::cout << "Output pipeline result matrix:\n";
266 q_res_output.print(std::cout);
267
268 // Expected result
269 std::cout << "Expected result:\n";
270 q_dst0.print(std::cout);
271 #endif // ARM_COMPUTE_DEBUG_ENABLED
272 }
273