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_leaky_relu_nc_q8(size_t channels,float negative_slope,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 * leaky_relu_out)19 enum pytorch_qnnp_status pytorch_qnnp_create_leaky_relu_nc_q8(
20 size_t channels,
21 float negative_slope,
22 uint8_t input_zero_point,
23 float input_scale,
24 uint8_t output_zero_point,
25 float output_scale,
26 uint8_t output_min,
27 uint8_t output_max,
28 uint32_t flags,
29 pytorch_qnnp_operator_t* leaky_relu_out) {
30 pytorch_qnnp_operator_t leaky_relu_op = NULL;
31 enum pytorch_qnnp_status status = pytorch_qnnp_status_uninitialized;
32
33 if (!pytorch_qnnp_params.initialized) {
34 pytorch_qnnp_log_error(
35 "pytorch_qnnp_create_leaky_relu_nc_q8 failed because QNNPACK is not properly initialized");
36 goto error;
37 }
38
39 status = pytorch_qnnp_status_invalid_parameter;
40
41 if (channels == 0) {
42 pytorch_qnnp_log_error(
43 "failed to create Leaky ReLU operator with %zu channels: number of channels must be non-zero",
44 channels);
45 goto error;
46 }
47
48 if (negative_slope <= 0.0f || !isnormal(negative_slope)) {
49 pytorch_qnnp_log_error(
50 "failed to create Leaky ReLU operator with %.7g negative slope: slope must be finite and positive",
51 negative_slope);
52 goto error;
53 }
54
55 if (negative_slope > 1.0f) {
56 pytorch_qnnp_log_error(
57 "failed to create Leaky ReLU operator with %.7g negative slope: slope must not exceed 1.0",
58 negative_slope);
59 goto error;
60 }
61
62 if (input_scale <= 0.0f || !isnormal(input_scale)) {
63 pytorch_qnnp_log_error(
64 "failed to create Leaky ReLU operator with %.7g input scale: scale must be finite and positive",
65 input_scale);
66 goto error;
67 }
68
69 if (output_scale <= 0.0f || !isnormal(output_scale)) {
70 pytorch_qnnp_log_error(
71 "failed to create Leaky ReLU operator with %.7g output scale: scale must be finite and positive",
72 output_scale);
73 goto error;
74 }
75
76 if (output_min >= output_max) {
77 pytorch_qnnp_log_error(
78 "failed to create Leaky ReLU operator with [%" PRIu8 ", %" PRIu8
79 "] output range: range min must be below range max",
80 output_min,
81 output_max);
82 goto error;
83 }
84
85 status = pytorch_qnnp_status_unsupported_parameter;
86
87 const float input_output_scale = input_scale / output_scale;
88 if (input_output_scale < 0x1.0p-8f || input_output_scale >= 0x1.0p+8f) {
89 pytorch_qnnp_log_error(
90 "failed to create Leaky ReLU operator with %.7g input-to-output scale ratio: "
91 "scale ratio must be in [2**-8, 2**8) range",
92 input_output_scale);
93 goto error;
94 }
95
96 status = pytorch_qnnp_status_out_of_memory;
97
98 leaky_relu_op = calloc(1, sizeof(struct pytorch_qnnp_operator));
99 if (leaky_relu_op == NULL) {
100 pytorch_qnnp_log_error(
101 "failed to allocate %zu bytes for pytorch_qnnp_operator structure",
102 sizeof(struct pytorch_qnnp_operator));
103 goto error;
104 }
105
106 leaky_relu_op->lookup_table = malloc(256 * sizeof(uint8_t));
107 if (leaky_relu_op->lookup_table == NULL) {
108 pytorch_qnnp_log_error(
109 "failed to allocate 256 bytes for Leaky ReLU lookup table");
110 goto error;
111 }
112
113 uint8_t* lookup_table = leaky_relu_op->lookup_table;
114 const float scaled_min_less_zero_point =
115 (float)((int32_t)output_min - (int32_t)output_zero_point);
116 const float scaled_max_less_zero_point =
117 (float)((int32_t)output_max - (int32_t)output_zero_point);
118 for (int32_t i = 0; i < 256; i++) {
119 const float x =
120 input_output_scale * (float)(i - (int32_t)(uint32_t)input_zero_point);
121 float y = x < 0.0f ? x * negative_slope : x;
122 if (y < scaled_min_less_zero_point) {
123 y = scaled_min_less_zero_point;
124 }
125 if (y > scaled_max_less_zero_point) {
126 y = scaled_max_less_zero_point;
127 }
128 lookup_table[(uint32_t)i] = (uint8_t)(lrintf(y) + (long)output_zero_point);
129 }
130
131 leaky_relu_op->channels = channels;
132
133 leaky_relu_op->ukernel_type = pytorch_qnnp_ukernel_type_lut;
134 leaky_relu_op->format = pytorch_qnnp_format_quint8;
135
136 *leaky_relu_out = leaky_relu_op;
137 return pytorch_qnnp_status_success;
138
139 error:
140 pytorch_qnnp_delete_operator(leaky_relu_op);
141 return status;
142 }
143
pytorch_qnnp_setup_leaky_relu_nc_q8(pytorch_qnnp_operator_t leaky_relu,size_t batch_size,const uint8_t * input,size_t input_stride,uint8_t * output,size_t output_stride)144 enum pytorch_qnnp_status pytorch_qnnp_setup_leaky_relu_nc_q8(
145 pytorch_qnnp_operator_t leaky_relu,
146 size_t batch_size,
147 const uint8_t* input,
148 size_t input_stride,
149 uint8_t* output,
150 size_t output_stride) {
151 if (!pytorch_qnnp_params.initialized) {
152 pytorch_qnnp_log_error(
153 "pytorch_qnnp_setup_leaky_relu_nc_q8 failed because QNNPACK is not properly initialized");
154 return pytorch_qnnp_status_uninitialized;
155 }
156
157 if (batch_size == 0) {
158 leaky_relu->batch_size = 0;
159 return pytorch_qnnp_status_success;
160 }
161
162 leaky_relu->batch_size = batch_size;
163 leaky_relu->input = input;
164 leaky_relu->input_pixel_stride = input_stride;
165 leaky_relu->output = output;
166 leaky_relu->output_pixel_stride = output_stride;
167
168 return pytorch_qnnp_status_success;
169 }
170