1 /*
2 * SPDX-FileCopyrightText: Copyright 2010-2024 Arm Limited and/or its affiliates <open-source-office@arm.com>
3 *
4 * SPDX-License-Identifier: Apache-2.0
5 *
6 * Licensed under the Apache License, Version 2.0 (the License); you may
7 * not use this file except in compliance with the License.
8 * You may obtain a copy of the License at
9 *
10 * www.apache.org/licenses/LICENSE-2.0
11 *
12 * Unless required by applicable law or agreed to in writing, software
13 * distributed under the License is distributed on an AS IS BASIS, WITHOUT
14 * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
15 * See the License for the specific language governing permissions and
16 * limitations under the License.
17 */
18
19 /* ----------------------------------------------------------------------
20 * Project: CMSIS NN Library
21 * Title: arm_svdf_state_s16_s8.c
22 * Description: S8 basic SVDF layer function with s16 state tensor
23 *
24 * $Date: 24 Sep 2024
25 * $Revision: V.3.1.1
26 *
27 * Target : Arm(R) M-Profile Architecture
28 *
29 * -------------------------------------------------------------------- */
30
31 #include "arm_nnfunctions.h"
32 #include "arm_nnsupportfunctions.h"
33
34 /**
35 * @ingroup Public
36 */
37
38 /**
39 * @addtogroup SVDF
40 * @{
41 */
42
43 /*
44 * S8 SVDF layer function for TensorFlow Lite with 16 bit state tensor
45 *
46 * Refer to header file for details.
47 *
48 */
49
arm_svdf_state_s16_s8(const cmsis_nn_context * input_ctx,const cmsis_nn_context * output_ctx,const cmsis_nn_svdf_params * svdf_params,const cmsis_nn_per_tensor_quant_params * input_quant_params,const cmsis_nn_per_tensor_quant_params * output_quant_params,const cmsis_nn_dims * input_dims,const int8_t * input_data,const cmsis_nn_dims * state_dims,int16_t * state_data,const cmsis_nn_dims * weights_feature_dims,const int8_t * weights_feature_data,const cmsis_nn_dims * weights_time_dims,const int16_t * weights_time_data,const cmsis_nn_dims * bias_dims,const int32_t * bias_data,const cmsis_nn_dims * output_dims,int8_t * output_data)50 arm_cmsis_nn_status arm_svdf_state_s16_s8(const cmsis_nn_context *input_ctx,
51 const cmsis_nn_context *output_ctx,
52 const cmsis_nn_svdf_params *svdf_params,
53 const cmsis_nn_per_tensor_quant_params *input_quant_params,
54 const cmsis_nn_per_tensor_quant_params *output_quant_params,
55 const cmsis_nn_dims *input_dims,
56 const int8_t *input_data,
57 const cmsis_nn_dims *state_dims,
58 int16_t *state_data,
59 const cmsis_nn_dims *weights_feature_dims,
60 const int8_t *weights_feature_data,
61 const cmsis_nn_dims *weights_time_dims,
62 const int16_t *weights_time_data,
63 const cmsis_nn_dims *bias_dims,
64 const int32_t *bias_data,
65 const cmsis_nn_dims *output_dims,
66 int8_t *output_data)
67 {
68 (void)bias_dims;
69 (void)state_dims;
70 (void)output_dims;
71
72 const int32_t multiplier_in = input_quant_params->multiplier;
73 const int32_t shift_in = input_quant_params->shift;
74 const int32_t multiplier_out = output_quant_params->multiplier;
75 const int32_t shift_2 = output_quant_params->shift;
76 const int32_t zp_in = svdf_params->input_offset;
77 const int32_t zp_out = svdf_params->output_offset;
78 const int32_t in_activation_min = svdf_params->input_activation.min;
79 const int32_t in_activation_max = svdf_params->input_activation.max;
80 const int32_t out_activation_min = svdf_params->output_activation.min;
81 const int32_t out_activation_max = svdf_params->output_activation.max;
82 const int16_t rank = svdf_params->rank;
83
84 const int32_t input_batches = input_dims->n;
85 const int32_t input_height = input_dims->h;
86 const int32_t feature_batches = weights_feature_dims->n;
87 const int32_t time_batches = weights_time_dims->h;
88 const int32_t unit_count = feature_batches / rank;
89
90 if (input_ctx->buf == NULL)
91 {
92 return ARM_CMSIS_NN_ARG_ERROR;
93 }
94 int32_t *buffer_a = (int32_t *)input_ctx->buf;
95
96 if (output_ctx->buf == NULL)
97 {
98 return ARM_CMSIS_NN_ARG_ERROR;
99 }
100 int32_t *buffer_b = (int32_t *)output_ctx->buf;
101
102 // Left shift state
103 // Using memcpy on overlapping data is in general undefined behaviour, but since the behaviour of arm_memcpy_s8 is
104 // known it is certain that the data has been copied before it is overwritten in this case.
105 #ifdef ARM_MATH_MVEI
106 arm_memcpy_s8((int8_t *)state_data,
107 (int8_t *)(state_data + 1),
108 (size_t)((input_batches * feature_batches * time_batches - 1) * (int32_t)sizeof(int16_t)));
109 #else
110 memmove(state_data,
111 state_data + 1,
112 (size_t)((input_batches * feature_batches * time_batches - 1) * (int32_t)sizeof(int16_t)));
113 #endif
114
115 // Matrix multiplication input * feature weight
116 for (int i_batch = 0; i_batch < input_batches; i_batch++)
117 {
118 int16_t *res_ptr = state_data + (time_batches * i_batch * feature_batches) + (time_batches - 1);
119 const int8_t *weight = weights_feature_data;
120 const int8_t *input = input_data + i_batch * input_height;
121
122 arm_cmsis_nn_status res = arm_nn_vec_mat_mult_t_svdf_s8(input,
123 weight,
124 res_ptr,
125 -zp_in,
126 time_batches,
127 multiplier_in,
128 shift_in,
129 input_height,
130 feature_batches,
131 in_activation_min,
132 in_activation_max);
133
134 if (res != ARM_CMSIS_NN_SUCCESS)
135 {
136 return res;
137 }
138 }
139
140 {
141 // Matrix multiplication time weight * state tensors
142 int32_t *ptr_a = buffer_a;
143 const int16_t *v2 = state_data;
144 for (int i_batch = 0; i_batch < input_batches; i_batch++)
145 {
146 const int16_t *v1 = weights_time_data;
147
148 for (int i_feature_batch = 0; i_feature_batch < feature_batches; i_feature_batch++)
149 {
150 *ptr_a = 0;
151 int32_t sum = 0;
152 #if defined(ARM_MATH_DSP) && !defined(ARM_MATH_MVEI)
153 // Perform matrix multiplication in blocks of two
154 int j = 0;
155 int32_t block_count = time_batches >> 1;
156 for (int i = 0; i < block_count; i++)
157 {
158 j += 2;
159 int32_t r1 = arm_nn_read_q15x2_ia(&v1);
160 int32_t r2 = arm_nn_read_q15x2_ia(&v2);
161
162 sum = SMLAD(r1, r2, sum);
163 }
164
165 // Process the remaining data
166 for (; j < time_batches; j++)
167 {
168 sum += *v1 * *v2;
169 v1++;
170 v2++;
171 }
172 #else
173 for (int j = 0; j < time_batches; j++)
174 {
175 sum += *v1 * *v2;
176 v1++;
177 v2++;
178 }
179 #endif
180
181 *ptr_a = sum;
182 ptr_a++;
183 }
184 }
185 }
186
187 if (bias_data)
188 {
189 if (unit_count == feature_batches)
190 {
191 for (int i = 0; i < input_batches; i++)
192 {
193 int32_t *output_temp = buffer_b + i * feature_batches;
194 const int32_t *ptr_a = buffer_a + i * feature_batches;
195
196 const int32_t *bi = bias_data;
197 for (int j = 0; j < feature_batches; j++)
198 {
199 output_temp[j] = ptr_a[j] + bi[j];
200 }
201 }
202 }
203 else
204 {
205 for (int i_batch = 0; i_batch < input_batches; i_batch++)
206 {
207 int32_t *output_data_temp = buffer_b + i_batch * unit_count;
208 int32_t *ptr_a = buffer_a + i_batch * feature_batches;
209
210 for (int i = 0; i < unit_count; i++)
211 {
212 int32_t sum = bias_data[i];
213 for (int j = 0; j < rank; j++)
214 {
215 sum += *ptr_a;
216 ptr_a++;
217 }
218 output_data_temp[i] = sum;
219 }
220 }
221 }
222 }
223 else
224 {
225 for (int i_batch = 0; i_batch < input_batches; i_batch++)
226 {
227 int32_t *output_data_temp = buffer_b + i_batch * unit_count;
228 int32_t *ptr_a = buffer_a + i_batch * feature_batches;
229
230 for (int i = 0; i < unit_count; i++)
231 {
232 int32_t sum = 0;
233 for (int j = 0; j < rank; j++)
234 {
235 sum += *ptr_a;
236 ptr_a++;
237 }
238 output_data_temp[i] = sum;
239 }
240 }
241 }
242
243 #if defined(ARM_MATH_MVEI)
244 int32_t num_elements = input_batches * unit_count;
245 const int32_t loop_count = (num_elements + 3) / 4;
246 for (int i_op = 0; i_op < loop_count; i_op++)
247 {
248 mve_pred16_t p = vctp32q((uint32_t)num_elements);
249 int32x4_t op = vldrwq_z_s32(buffer_b, p);
250 op = arm_requantize_mve(op, multiplier_out, shift_2);
251 op = vaddq_n_s32(op, zp_out);
252 const int32x4_t min_vec = vdupq_n_s32((int8_t)out_activation_min);
253 const int32x4_t max_vec = vdupq_n_s32((int8_t)out_activation_max);
254 op = vmaxq_s32(op, min_vec);
255 op = vminq_s32(op, max_vec);
256 vstrbq_p_s32(output_data, op, p);
257 output_data += 4;
258 buffer_b += 4;
259 num_elements -= 4;
260 }
261 #else
262 for (int i = 0; i < input_batches * unit_count; i++)
263 {
264 output_data[i] = (int8_t)CLAMP(
265 arm_nn_requantize(buffer_b[i], multiplier_out, shift_2) + zp_out, out_activation_max, out_activation_min);
266 }
267 #endif
268
269 return (ARM_CMSIS_NN_SUCCESS);
270 }
271
272 /**
273 * @} end of SVDF group
274 */
275
