1 /*
2 * Copyright (c) Facebook, Inc. and its affiliates.
3 * All rights reserved.
4 *
5 * This source code is licensed under the BSD-style license found in the
6 * LICENSE file in the root directory of this source tree.
7 */
8
9 #include <assert.h>
10 #include <math.h>
11 #include <stddef.h>
12 #include <stdint.h>
13 #include <stdlib.h>
14
15 #include <pytorch_qnnpack.h>
16 #include <qnnpack/log.h>
17 #include <qnnpack/operator.h>
18
pytorch_qnnp_create_hardsigmoid_nc_q8(size_t channels,uint8_t input_zero_point,float input_scale,uint8_t output_zero_point,float output_scale,uint8_t output_min,uint8_t output_max,uint32_t flags,pytorch_qnnp_operator_t * hardsigmoid_out)19 enum pytorch_qnnp_status pytorch_qnnp_create_hardsigmoid_nc_q8(
20 size_t channels,
21 uint8_t input_zero_point,
22 float input_scale,
23 uint8_t output_zero_point,
24 float output_scale,
25 uint8_t output_min,
26 uint8_t output_max,
27 uint32_t flags,
28 pytorch_qnnp_operator_t* hardsigmoid_out) {
29 pytorch_qnnp_operator_t hardsigmoid_op = NULL;
30 enum pytorch_qnnp_status status = pytorch_qnnp_status_uninitialized;
31
32 if (!pytorch_qnnp_params.initialized) {
33 pytorch_qnnp_log_error(
34 "pytorch_qnnp_create_hardsigmoid_nc_q8 failed because QNNPACK is not properly initialized");
35 goto error;
36 }
37
38 status = pytorch_qnnp_status_invalid_parameter;
39
40 if (channels == 0) {
41 pytorch_qnnp_log_error(
42 "failed to create Hardsigmoid operator with %zu channels: number of channels must be non-zero",
43 channels);
44 goto error;
45 }
46
47 if (input_scale <= 0.0f || !isnormal(input_scale)) {
48 pytorch_qnnp_log_error(
49 "failed to create Hardsigmoid operator with %.7g input scale: scale must be finite and positive",
50 input_scale);
51 goto error;
52 }
53
54 if (output_scale <= 0.0f || !isnormal(output_scale)) {
55 pytorch_qnnp_log_error(
56 "failed to create Hardsigmoid operator with %.7g output scale: scale must be finite and positive",
57 output_scale);
58 goto error;
59 }
60
61 if (output_min >= output_max) {
62 pytorch_qnnp_log_error(
63 "failed to create Hardsigmoid operator with [%" PRIu8 ", %" PRIu8
64 "] output range: range min must be below range max",
65 output_min,
66 output_max);
67 goto error;
68 }
69
70 status = pytorch_qnnp_status_unsupported_parameter;
71
72 if (output_scale != 0x1.0p-8f) {
73 pytorch_qnnp_log_error(
74 "failed to create Hardsigmoid operator with %.7g output scale: only output scale of 1/256 is supported",
75 output_scale);
76 goto error;
77 }
78
79 if (output_zero_point != 0) {
80 pytorch_qnnp_log_error(
81 "failed to create Hardsigmoid operator with %" PRIu8
82 " output zero point: only output zero point of 0 is supported",
83 output_zero_point);
84 goto error;
85 }
86
87 status = pytorch_qnnp_status_out_of_memory;
88
89 hardsigmoid_op = calloc(1, sizeof(struct pytorch_qnnp_operator));
90 if (hardsigmoid_op == NULL) {
91 pytorch_qnnp_log_error(
92 "failed to allocate %zu bytes for pytorch_qnnp_operator structure",
93 sizeof(struct pytorch_qnnp_operator));
94 goto error;
95 }
96
97 hardsigmoid_op->lookup_table = malloc(256 * sizeof(uint8_t));
98 if (hardsigmoid_op->lookup_table == NULL) {
99 pytorch_qnnp_log_error(
100 "failed to allocate 256 bytes for Hardsigmoid lookup table");
101 goto error;
102 }
103
104 uint8_t* lookup_table = hardsigmoid_op->lookup_table;
105 const float scaled_min = (float)(int32_t)output_min;
106 const float scaled_max = (float)(int32_t)output_max;
107 const float inv_output_scale = 1.0f / output_scale;
108 for (int32_t i = 0; i < 256; i++) {
109 float x =
110 input_scale * (float)(i - (int32_t)(uint32_t)input_zero_point);
111 // hardsigmoid, no min/max functions in C
112 float x2 = x + 3.0f;
113 x2 = x2 > 0.0f ? x2 : 0.0f;
114 x2 = x2 < 6.0f ? x2 : 6.0f;
115 x2 = x2 / 6.0f;
116 float scaled_hardsigmoid_x = inv_output_scale * x2 + output_zero_point;
117 if (scaled_hardsigmoid_x < scaled_min) {
118 scaled_hardsigmoid_x = scaled_min;
119 }
120 if (scaled_hardsigmoid_x > scaled_max) {
121 scaled_hardsigmoid_x = scaled_max;
122 }
123 lookup_table[(uint32_t)i] = (uint8_t)lrintf(scaled_hardsigmoid_x);
124 }
125
126 hardsigmoid_op->channels = channels;
127
128 hardsigmoid_op->ukernel_type = pytorch_qnnp_ukernel_type_lut;
129 hardsigmoid_op->format = pytorch_qnnp_format_quint8;
130
131 *hardsigmoid_out = hardsigmoid_op;
132 return pytorch_qnnp_status_success;
133
134 error:
135 pytorch_qnnp_delete_operator(hardsigmoid_op);
136 return status;
137 }
138
pytorch_qnnp_setup_hardsigmoid_nc_q8(pytorch_qnnp_operator_t hardsigmoid,size_t batch_size,const uint8_t * input,size_t input_stride,uint8_t * output,size_t output_stride)139 enum pytorch_qnnp_status pytorch_qnnp_setup_hardsigmoid_nc_q8(
140 pytorch_qnnp_operator_t hardsigmoid,
141 size_t batch_size,
142 const uint8_t* input,
143 size_t input_stride,
144 uint8_t* output,
145 size_t output_stride) {
146 if (!pytorch_qnnp_params.initialized) {
147 pytorch_qnnp_log_error(
148 "pytorch_qnnp_setup_hardsigmoid_nc_q8 failed because QNNPACK is not properly initialized");
149 return pytorch_qnnp_status_uninitialized;
150 }
151
152 if (batch_size == 0) {
153 hardsigmoid->batch_size = 0;
154 return pytorch_qnnp_status_success;
155 }
156
157 hardsigmoid->batch_size = batch_size;
158 hardsigmoid->input = input;
159 hardsigmoid->input_pixel_stride = input_stride;
160 hardsigmoid->output = output;
161 hardsigmoid->output_pixel_stride = output_stride;
162
163 return pytorch_qnnp_status_success;
164 }
165