1 /*
2 * SPDX-FileCopyrightText: Copyright 2010-2023 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_s8.c
22 * Description: S8 basic SVDF layer function with s16 state tensor
23 *
24 * $Date: 5 January 2023
25 * $Revision: V.3.1.0
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 memmove((int16_t *)state_data,
104 (int16_t *)state_data + 1,
105 (size_t)((input_batches * feature_batches * time_batches - 1) * (int32_t)sizeof(int16_t)));
106
107 // Matrix multiplication input * feature weight
108 for (int i_batch = 0; i_batch < input_batches; i_batch++)
109 {
110 int16_t *res_ptr = state_data + (time_batches * i_batch * feature_batches) + (time_batches - 1);
111 const int8_t *weight = weights_feature_data;
112 const int8_t *input = input_data + i_batch * input_height;
113
114 arm_cmsis_nn_status res = arm_nn_vec_mat_mult_t_svdf_s8(input,
115 weight,
116 res_ptr,
117 -zp_in,
118 time_batches,
119 multiplier_in,
120 shift_in,
121 input_height,
122 feature_batches,
123 in_activation_min,
124 in_activation_max);
125
126 if (res != ARM_CMSIS_NN_SUCCESS)
127 {
128 return res;
129 }
130 }
131
132 {
133 // Matrix multiplication time weight * state tensors
134 int32_t *ptr_a = buffer_a;
135 const int16_t *v2 = state_data;
136 for (int i_batch = 0; i_batch < input_batches; i_batch++)
137 {
138 const int16_t *v1 = weights_time_data;
139
140 for (int i_feature_batch = 0; i_feature_batch < feature_batches; i_feature_batch++)
141 {
142 *ptr_a = 0;
143 int32_t sum = 0;
144 #if defined(ARM_MATH_DSP) && !defined(ARM_MATH_MVEI)
145 // Perform matrix multiplication in blocks of two
146 int j = 0;
147 int32_t block_count = time_batches >> 1;
148 for (int i = 0; i < block_count; i++)
149 {
150 j += 2;
151 int32_t r1 = arm_nn_read_q15x2_ia(&v1);
152 int32_t r2 = arm_nn_read_q15x2_ia(&v2);
153
154 sum = SMLAD(r1, r2, sum);
155 }
156
157 // Process the remaining data
158 for (; j < time_batches; j++)
159 {
160 sum += *v1 * *v2;
161 v1++;
162 v2++;
163 }
164 #else
165 for (int j = 0; j < time_batches; j++)
166 {
167 sum += *v1 * *v2;
168 v1++;
169 v2++;
170 }
171 #endif
172
173 *ptr_a = sum;
174 ptr_a++;
175 }
176 }
177 }
178
179 if (bias_data)
180 {
181 if (unit_count == feature_batches)
182 {
183 for (int i = 0; i < input_batches; i++)
184 {
185 int32_t *output_temp = buffer_b + i * feature_batches;
186 const int32_t *ptr_a = buffer_a + i * feature_batches;
187
188 const int32_t *bi = bias_data;
189 for (int j = 0; j < feature_batches; j++)
190 {
191 output_temp[j] = ptr_a[j] + bi[j];
192 }
193 }
194 }
195 else
196 {
197 for (int i_batch = 0; i_batch < input_batches; i_batch++)
198 {
199 int32_t *output_data_temp = buffer_b + i_batch * unit_count;
200 int32_t *ptr_a = buffer_a + i_batch * feature_batches;
201
202 for (int i = 0; i < unit_count; i++)
203 {
204 int32_t sum = bias_data[i];
205 for (int j = 0; j < rank; j++)
206 {
207 sum += *ptr_a;
208 ptr_a++;
209 }
210 output_data_temp[i] = sum;
211 }
212 }
213 }
214 }
215 else
216 {
217 for (int i_batch = 0; i_batch < input_batches; i_batch++)
218 {
219 int32_t *output_data_temp = buffer_b + i_batch * unit_count;
220 int32_t *ptr_a = buffer_a + i_batch * feature_batches;
221
222 for (int i = 0; i < unit_count; i++)
223 {
224 int32_t sum = 0;
225 for (int j = 0; j < rank; j++)
226 {
227 sum += *ptr_a;
228 ptr_a++;
229 }
230 output_data_temp[i] = sum;
231 }
232 }
233 }
234
235 #if defined(ARM_MATH_MVEI)
236 int32_t num_elements = input_batches * unit_count;
237 const int32_t loop_count = (num_elements + 3) / 4;
238 for (int i_op = 0; i_op < loop_count; i_op++)
239 {
240 mve_pred16_t p = vctp32q((uint32_t)num_elements);
241 int32x4_t op = vldrwq_z_s32(buffer_b, p);
242 op = arm_requantize_mve(op, multiplier_out, shift_2);
243 op = vaddq_n_s32(op, zp_out);
244 const int32x4_t min_vec = vdupq_n_s32((int8_t)out_activation_min);
245 const int32x4_t max_vec = vdupq_n_s32((int8_t)out_activation_max);
246 op = vmaxq_s32(op, min_vec);
247 op = vminq_s32(op, max_vec);
248 vstrbq_p_s32(output_data, op, p);
249 output_data += 4;
250 buffer_b += 4;
251 num_elements -= 4;
252 }
253 #else
254 for (int i = 0; i < input_batches * unit_count; i++)
255 {
256 output_data[i] = (int8_t)CLAMP(
257 arm_nn_requantize(buffer_b[i], multiplier_out, shift_2) + zp_out, out_activation_max, out_activation_min);
258 }
259 #endif
260
261 return (ARM_CMSIS_NN_SUCCESS);
262 }
263
264 /**
265 * @} end of SVDF group
266 */
267
