xref: /aosp_15_r20/external/XNNPACK/src/operators/channel-shuffle-nc.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Copyright (c) Facebook, Inc. and its affiliates.
2*4bdc9457SAndroid Build Coastguard Worker // All rights reserved.
3*4bdc9457SAndroid Build Coastguard Worker //
4*4bdc9457SAndroid Build Coastguard Worker // Copyright 2019 Google LLC
5*4bdc9457SAndroid Build Coastguard Worker //
6*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
7*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
8*4bdc9457SAndroid Build Coastguard Worker 
9*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
10*4bdc9457SAndroid Build Coastguard Worker #include <math.h>
11*4bdc9457SAndroid Build Coastguard Worker #include <stddef.h>
12*4bdc9457SAndroid Build Coastguard Worker #include <stdint.h>
13*4bdc9457SAndroid Build Coastguard Worker #include <stdlib.h>
14*4bdc9457SAndroid Build Coastguard Worker 
15*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack.h>
16*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/allocator.h>
17*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/operator.h>
18*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/log.h>
19*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/params.h>
20*4bdc9457SAndroid Build Coastguard Worker 
21*4bdc9457SAndroid Build Coastguard Worker 
create_channel_shuffle_nc(size_t groups,size_t group_channels,size_t input_stride,size_t output_stride,uint32_t flags,enum xnn_operator_type operator_type,xnn_operator_t * channel_shuffle_op_out)22*4bdc9457SAndroid Build Coastguard Worker static enum xnn_status create_channel_shuffle_nc(
23*4bdc9457SAndroid Build Coastguard Worker   size_t groups,
24*4bdc9457SAndroid Build Coastguard Worker   size_t group_channels,
25*4bdc9457SAndroid Build Coastguard Worker   size_t input_stride,
26*4bdc9457SAndroid Build Coastguard Worker   size_t output_stride,
27*4bdc9457SAndroid Build Coastguard Worker   uint32_t flags,
28*4bdc9457SAndroid Build Coastguard Worker   enum xnn_operator_type operator_type,
29*4bdc9457SAndroid Build Coastguard Worker   xnn_operator_t* channel_shuffle_op_out)
30*4bdc9457SAndroid Build Coastguard Worker {
31*4bdc9457SAndroid Build Coastguard Worker   xnn_operator_t channel_shuffle_op = NULL;
32*4bdc9457SAndroid Build Coastguard Worker   enum xnn_status status = xnn_status_uninitialized;
33*4bdc9457SAndroid Build Coastguard Worker 
34*4bdc9457SAndroid Build Coastguard Worker   if ((xnn_params.init_flags & XNN_INIT_FLAG_XNNPACK) == 0) {
35*4bdc9457SAndroid Build Coastguard Worker     xnn_log_error("failed to create %s operator: XNNPACK is not initialized",
36*4bdc9457SAndroid Build Coastguard Worker       xnn_operator_type_to_string(operator_type));
37*4bdc9457SAndroid Build Coastguard Worker     goto error;
38*4bdc9457SAndroid Build Coastguard Worker   }
39*4bdc9457SAndroid Build Coastguard Worker 
40*4bdc9457SAndroid Build Coastguard Worker   status = xnn_status_invalid_parameter;
41*4bdc9457SAndroid Build Coastguard Worker 
42*4bdc9457SAndroid Build Coastguard Worker   if (groups <= 1) {
43*4bdc9457SAndroid Build Coastguard Worker     xnn_log_error(
44*4bdc9457SAndroid Build Coastguard Worker       "failed to create %s operator with %zu groups: at least two groups required",
45*4bdc9457SAndroid Build Coastguard Worker       xnn_operator_type_to_string(operator_type), groups);
46*4bdc9457SAndroid Build Coastguard Worker     goto error;
47*4bdc9457SAndroid Build Coastguard Worker   }
48*4bdc9457SAndroid Build Coastguard Worker 
49*4bdc9457SAndroid Build Coastguard Worker   if (group_channels == 0) {
50*4bdc9457SAndroid Build Coastguard Worker     xnn_log_error(
51*4bdc9457SAndroid Build Coastguard Worker       "failed to create %s operator with %zu group channels: number of group channels must be non-zero",
52*4bdc9457SAndroid Build Coastguard Worker       xnn_operator_type_to_string(operator_type), group_channels);
53*4bdc9457SAndroid Build Coastguard Worker     goto error;
54*4bdc9457SAndroid Build Coastguard Worker   }
55*4bdc9457SAndroid Build Coastguard Worker 
56*4bdc9457SAndroid Build Coastguard Worker   const size_t channels = groups * group_channels;
57*4bdc9457SAndroid Build Coastguard Worker   if (input_stride < channels) {
58*4bdc9457SAndroid Build Coastguard Worker     xnn_log_error(
59*4bdc9457SAndroid Build Coastguard Worker       "failed to create %s operator with input element stride of %zu: "
60*4bdc9457SAndroid Build Coastguard Worker       "stride must be at least as large as the number of channels (%zux%zu)",
61*4bdc9457SAndroid Build Coastguard Worker       xnn_operator_type_to_string(operator_type), input_stride, groups, group_channels);
62*4bdc9457SAndroid Build Coastguard Worker     goto error;
63*4bdc9457SAndroid Build Coastguard Worker   }
64*4bdc9457SAndroid Build Coastguard Worker 
65*4bdc9457SAndroid Build Coastguard Worker   if (output_stride < channels) {
66*4bdc9457SAndroid Build Coastguard Worker     xnn_log_error(
67*4bdc9457SAndroid Build Coastguard Worker       "failed to create %s operator with output element stride of %zu: "
68*4bdc9457SAndroid Build Coastguard Worker       "stride must be at least as large as the number of channels (%zux%zu)",
69*4bdc9457SAndroid Build Coastguard Worker       xnn_operator_type_to_string(operator_type), output_stride, groups, group_channels);
70*4bdc9457SAndroid Build Coastguard Worker     goto error;
71*4bdc9457SAndroid Build Coastguard Worker   }
72*4bdc9457SAndroid Build Coastguard Worker 
73*4bdc9457SAndroid Build Coastguard Worker   status = xnn_status_out_of_memory;
74*4bdc9457SAndroid Build Coastguard Worker 
75*4bdc9457SAndroid Build Coastguard Worker   channel_shuffle_op = xnn_allocate_zero_simd_memory(sizeof(struct xnn_operator));
76*4bdc9457SAndroid Build Coastguard Worker   if (channel_shuffle_op == NULL) {
77*4bdc9457SAndroid Build Coastguard Worker     xnn_log_error(
78*4bdc9457SAndroid Build Coastguard Worker       "failed to allocate %zu bytes for %s operator descriptor",
79*4bdc9457SAndroid Build Coastguard Worker       sizeof(struct xnn_operator), xnn_operator_type_to_string(operator_type));
80*4bdc9457SAndroid Build Coastguard Worker     goto error;
81*4bdc9457SAndroid Build Coastguard Worker   }
82*4bdc9457SAndroid Build Coastguard Worker 
83*4bdc9457SAndroid Build Coastguard Worker   channel_shuffle_op->groups = groups;
84*4bdc9457SAndroid Build Coastguard Worker   channel_shuffle_op->group_channels = group_channels;
85*4bdc9457SAndroid Build Coastguard Worker   channel_shuffle_op->input_pixel_stride = input_stride;
86*4bdc9457SAndroid Build Coastguard Worker   channel_shuffle_op->output_pixel_stride = output_stride;
87*4bdc9457SAndroid Build Coastguard Worker 
88*4bdc9457SAndroid Build Coastguard Worker   channel_shuffle_op->type = operator_type;
89*4bdc9457SAndroid Build Coastguard Worker   channel_shuffle_op->flags = flags;
90*4bdc9457SAndroid Build Coastguard Worker 
91*4bdc9457SAndroid Build Coastguard Worker   channel_shuffle_op->state = xnn_run_state_invalid;
92*4bdc9457SAndroid Build Coastguard Worker 
93*4bdc9457SAndroid Build Coastguard Worker   *channel_shuffle_op_out = channel_shuffle_op;
94*4bdc9457SAndroid Build Coastguard Worker   return xnn_status_success;
95*4bdc9457SAndroid Build Coastguard Worker 
96*4bdc9457SAndroid Build Coastguard Worker error:
97*4bdc9457SAndroid Build Coastguard Worker   xnn_delete_operator(channel_shuffle_op);
98*4bdc9457SAndroid Build Coastguard Worker   return status;
99*4bdc9457SAndroid Build Coastguard Worker }
100*4bdc9457SAndroid Build Coastguard Worker 
101*4bdc9457SAndroid Build Coastguard Worker 
xnn_create_channel_shuffle_nc_x8(size_t groups,size_t group_channels,size_t input_stride,size_t output_stride,uint32_t flags,xnn_operator_t * channel_shuffle_op_out)102*4bdc9457SAndroid Build Coastguard Worker enum xnn_status xnn_create_channel_shuffle_nc_x8(
103*4bdc9457SAndroid Build Coastguard Worker     size_t groups,
104*4bdc9457SAndroid Build Coastguard Worker     size_t group_channels,
105*4bdc9457SAndroid Build Coastguard Worker     size_t input_stride,
106*4bdc9457SAndroid Build Coastguard Worker     size_t output_stride,
107*4bdc9457SAndroid Build Coastguard Worker     uint32_t flags,
108*4bdc9457SAndroid Build Coastguard Worker     xnn_operator_t* channel_shuffle_op_out)
109*4bdc9457SAndroid Build Coastguard Worker {
110*4bdc9457SAndroid Build Coastguard Worker   return create_channel_shuffle_nc(
111*4bdc9457SAndroid Build Coastguard Worker     groups,
112*4bdc9457SAndroid Build Coastguard Worker     group_channels,
113*4bdc9457SAndroid Build Coastguard Worker     input_stride,
114*4bdc9457SAndroid Build Coastguard Worker     output_stride,
115*4bdc9457SAndroid Build Coastguard Worker     flags,
116*4bdc9457SAndroid Build Coastguard Worker     xnn_operator_type_channel_shuffle_nc_x8,
117*4bdc9457SAndroid Build Coastguard Worker     channel_shuffle_op_out);
118*4bdc9457SAndroid Build Coastguard Worker }
119*4bdc9457SAndroid Build Coastguard Worker 
xnn_create_channel_shuffle_nc_x32(size_t groups,size_t group_channels,size_t input_stride,size_t output_stride,uint32_t flags,xnn_operator_t * channel_shuffle_op_out)120*4bdc9457SAndroid Build Coastguard Worker enum xnn_status xnn_create_channel_shuffle_nc_x32(
121*4bdc9457SAndroid Build Coastguard Worker     size_t groups,
122*4bdc9457SAndroid Build Coastguard Worker     size_t group_channels,
123*4bdc9457SAndroid Build Coastguard Worker     size_t input_stride,
124*4bdc9457SAndroid Build Coastguard Worker     size_t output_stride,
125*4bdc9457SAndroid Build Coastguard Worker     uint32_t flags,
126*4bdc9457SAndroid Build Coastguard Worker     xnn_operator_t* channel_shuffle_op_out)
127*4bdc9457SAndroid Build Coastguard Worker {
128*4bdc9457SAndroid Build Coastguard Worker   return create_channel_shuffle_nc(
129*4bdc9457SAndroid Build Coastguard Worker     groups,
130*4bdc9457SAndroid Build Coastguard Worker     group_channels,
131*4bdc9457SAndroid Build Coastguard Worker     input_stride,
132*4bdc9457SAndroid Build Coastguard Worker     output_stride,
133*4bdc9457SAndroid Build Coastguard Worker     flags,
134*4bdc9457SAndroid Build Coastguard Worker     xnn_operator_type_channel_shuffle_nc_x32,
135*4bdc9457SAndroid Build Coastguard Worker     channel_shuffle_op_out);
136*4bdc9457SAndroid Build Coastguard Worker }
137*4bdc9457SAndroid Build Coastguard Worker 
setup_channel_shuffle_nc(xnn_operator_t channel_shuffle_op,size_t batch_size,const void * input,void * output,uint32_t log2_element_size,const struct zip_parameters zip[restrict XNN_MIN_ELEMENTS (1)])138*4bdc9457SAndroid Build Coastguard Worker static enum xnn_status setup_channel_shuffle_nc(
139*4bdc9457SAndroid Build Coastguard Worker     xnn_operator_t channel_shuffle_op,
140*4bdc9457SAndroid Build Coastguard Worker     size_t batch_size,
141*4bdc9457SAndroid Build Coastguard Worker     const void* input,
142*4bdc9457SAndroid Build Coastguard Worker     void* output,
143*4bdc9457SAndroid Build Coastguard Worker     uint32_t log2_element_size,
144*4bdc9457SAndroid Build Coastguard Worker     const struct zip_parameters zip[restrict XNN_MIN_ELEMENTS(1)])
145*4bdc9457SAndroid Build Coastguard Worker {
146*4bdc9457SAndroid Build Coastguard Worker   channel_shuffle_op->state = xnn_run_state_invalid;
147*4bdc9457SAndroid Build Coastguard Worker 
148*4bdc9457SAndroid Build Coastguard Worker   if ((xnn_params.init_flags & XNN_INIT_FLAG_XNNPACK) == 0) {
149*4bdc9457SAndroid Build Coastguard Worker     xnn_log_error("failed to setup %s operator: XNNPACK is not initialized",
150*4bdc9457SAndroid Build Coastguard Worker       xnn_operator_type_to_string(channel_shuffle_op->type));
151*4bdc9457SAndroid Build Coastguard Worker     return xnn_status_uninitialized;
152*4bdc9457SAndroid Build Coastguard Worker   }
153*4bdc9457SAndroid Build Coastguard Worker 
154*4bdc9457SAndroid Build Coastguard Worker   if (batch_size == 0) {
155*4bdc9457SAndroid Build Coastguard Worker     channel_shuffle_op->state = xnn_run_state_skip;
156*4bdc9457SAndroid Build Coastguard Worker     return xnn_status_success;
157*4bdc9457SAndroid Build Coastguard Worker   }
158*4bdc9457SAndroid Build Coastguard Worker 
159*4bdc9457SAndroid Build Coastguard Worker   channel_shuffle_op->batch_size = batch_size;
160*4bdc9457SAndroid Build Coastguard Worker   channel_shuffle_op->input = input;
161*4bdc9457SAndroid Build Coastguard Worker   channel_shuffle_op->output = output;
162*4bdc9457SAndroid Build Coastguard Worker 
163*4bdc9457SAndroid Build Coastguard Worker   const size_t groups = channel_shuffle_op->groups;
164*4bdc9457SAndroid Build Coastguard Worker   channel_shuffle_op->context.channel_shuffle = (struct channel_shuffle_context) {
165*4bdc9457SAndroid Build Coastguard Worker     .x = input,
166*4bdc9457SAndroid Build Coastguard Worker     .x_stride = channel_shuffle_op->input_pixel_stride << log2_element_size,
167*4bdc9457SAndroid Build Coastguard Worker     .y = output,
168*4bdc9457SAndroid Build Coastguard Worker     .y_stride = channel_shuffle_op->output_pixel_stride << log2_element_size,
169*4bdc9457SAndroid Build Coastguard Worker     .n = channel_shuffle_op->group_channels << log2_element_size,
170*4bdc9457SAndroid Build Coastguard Worker     .m = groups,
171*4bdc9457SAndroid Build Coastguard Worker   };
172*4bdc9457SAndroid Build Coastguard Worker   channel_shuffle_op->compute.type = xnn_parallelization_type_1d;
173*4bdc9457SAndroid Build Coastguard Worker   channel_shuffle_op->compute.range[0] = batch_size;
174*4bdc9457SAndroid Build Coastguard Worker   switch (groups) {
175*4bdc9457SAndroid Build Coastguard Worker     case 2:
176*4bdc9457SAndroid Build Coastguard Worker       channel_shuffle_op->compute.task_1d = (pthreadpool_task_1d_t) xnn_compute_channel_shuffle_fixed;
177*4bdc9457SAndroid Build Coastguard Worker       channel_shuffle_op->context.channel_shuffle.fixed_ukernel = zip->x2;
178*4bdc9457SAndroid Build Coastguard Worker       break;
179*4bdc9457SAndroid Build Coastguard Worker     case 3:
180*4bdc9457SAndroid Build Coastguard Worker       channel_shuffle_op->compute.task_1d = (pthreadpool_task_1d_t) xnn_compute_channel_shuffle_fixed;
181*4bdc9457SAndroid Build Coastguard Worker       channel_shuffle_op->context.channel_shuffle.fixed_ukernel = zip->x3;
182*4bdc9457SAndroid Build Coastguard Worker       break;
183*4bdc9457SAndroid Build Coastguard Worker     case 4:
184*4bdc9457SAndroid Build Coastguard Worker       channel_shuffle_op->compute.task_1d = (pthreadpool_task_1d_t) xnn_compute_channel_shuffle_fixed;
185*4bdc9457SAndroid Build Coastguard Worker       channel_shuffle_op->context.channel_shuffle.fixed_ukernel = zip->x4;
186*4bdc9457SAndroid Build Coastguard Worker       break;
187*4bdc9457SAndroid Build Coastguard Worker     default:
188*4bdc9457SAndroid Build Coastguard Worker       channel_shuffle_op->compute.task_1d = (pthreadpool_task_1d_t) xnn_compute_channel_shuffle_variable;
189*4bdc9457SAndroid Build Coastguard Worker       channel_shuffle_op->context.channel_shuffle.variable_ukernel = zip->xm;
190*4bdc9457SAndroid Build Coastguard Worker       break;
191*4bdc9457SAndroid Build Coastguard Worker     case 0:
192*4bdc9457SAndroid Build Coastguard Worker     case 1:
193*4bdc9457SAndroid Build Coastguard Worker       XNN_UNREACHABLE;
194*4bdc9457SAndroid Build Coastguard Worker   }
195*4bdc9457SAndroid Build Coastguard Worker   channel_shuffle_op->state = xnn_run_state_ready;
196*4bdc9457SAndroid Build Coastguard Worker 
197*4bdc9457SAndroid Build Coastguard Worker   return xnn_status_success;
198*4bdc9457SAndroid Build Coastguard Worker }
199*4bdc9457SAndroid Build Coastguard Worker 
xnn_setup_channel_shuffle_nc_x8(xnn_operator_t channel_shuffle_op,size_t batch_size,const void * input,void * output,pthreadpool_t threadpool)200*4bdc9457SAndroid Build Coastguard Worker enum xnn_status xnn_setup_channel_shuffle_nc_x8(
201*4bdc9457SAndroid Build Coastguard Worker     xnn_operator_t channel_shuffle_op,
202*4bdc9457SAndroid Build Coastguard Worker     size_t batch_size,
203*4bdc9457SAndroid Build Coastguard Worker     const void* input,
204*4bdc9457SAndroid Build Coastguard Worker     void* output,
205*4bdc9457SAndroid Build Coastguard Worker     pthreadpool_t threadpool)
206*4bdc9457SAndroid Build Coastguard Worker {
207*4bdc9457SAndroid Build Coastguard Worker   if (channel_shuffle_op->type != xnn_operator_type_channel_shuffle_nc_x8) {
208*4bdc9457SAndroid Build Coastguard Worker     xnn_log_error("failed to setup operator: operator type mismatch (expected %s, got %s)",
209*4bdc9457SAndroid Build Coastguard Worker       xnn_operator_type_to_string(xnn_operator_type_channel_shuffle_nc_x8),
210*4bdc9457SAndroid Build Coastguard Worker       xnn_operator_type_to_string(channel_shuffle_op->type));
211*4bdc9457SAndroid Build Coastguard Worker     return xnn_status_invalid_parameter;
212*4bdc9457SAndroid Build Coastguard Worker   }
213*4bdc9457SAndroid Build Coastguard Worker 
214*4bdc9457SAndroid Build Coastguard Worker   return setup_channel_shuffle_nc(
215*4bdc9457SAndroid Build Coastguard Worker     channel_shuffle_op,
216*4bdc9457SAndroid Build Coastguard Worker     batch_size,
217*4bdc9457SAndroid Build Coastguard Worker     input,
218*4bdc9457SAndroid Build Coastguard Worker     output,
219*4bdc9457SAndroid Build Coastguard Worker     0 /* log2(sizeof(element)) = log2(sizeof(uint8_t)) */,
220*4bdc9457SAndroid Build Coastguard Worker     &xnn_params.x8.zip);
221*4bdc9457SAndroid Build Coastguard Worker }
222*4bdc9457SAndroid Build Coastguard Worker 
xnn_setup_channel_shuffle_nc_x32(xnn_operator_t channel_shuffle_op,size_t batch_size,const void * input,void * output,pthreadpool_t threadpool)223*4bdc9457SAndroid Build Coastguard Worker enum xnn_status xnn_setup_channel_shuffle_nc_x32(
224*4bdc9457SAndroid Build Coastguard Worker     xnn_operator_t channel_shuffle_op,
225*4bdc9457SAndroid Build Coastguard Worker     size_t batch_size,
226*4bdc9457SAndroid Build Coastguard Worker     const void* input,
227*4bdc9457SAndroid Build Coastguard Worker     void* output,
228*4bdc9457SAndroid Build Coastguard Worker     pthreadpool_t threadpool)
229*4bdc9457SAndroid Build Coastguard Worker {
230*4bdc9457SAndroid Build Coastguard Worker   if (channel_shuffle_op->type != xnn_operator_type_channel_shuffle_nc_x32) {
231*4bdc9457SAndroid Build Coastguard Worker     xnn_log_error("failed to setup operator: operator type mismatch (expected %s, got %s)",
232*4bdc9457SAndroid Build Coastguard Worker       xnn_operator_type_to_string(xnn_operator_type_channel_shuffle_nc_x32),
233*4bdc9457SAndroid Build Coastguard Worker       xnn_operator_type_to_string(channel_shuffle_op->type));
234*4bdc9457SAndroid Build Coastguard Worker     return xnn_status_invalid_parameter;
235*4bdc9457SAndroid Build Coastguard Worker   }
236*4bdc9457SAndroid Build Coastguard Worker 
237*4bdc9457SAndroid Build Coastguard Worker   return setup_channel_shuffle_nc(
238*4bdc9457SAndroid Build Coastguard Worker     channel_shuffle_op,
239*4bdc9457SAndroid Build Coastguard Worker     batch_size,
240*4bdc9457SAndroid Build Coastguard Worker     input,
241*4bdc9457SAndroid Build Coastguard Worker     output,
242*4bdc9457SAndroid Build Coastguard Worker     2 /* log2(sizeof(element)) = log2(sizeof(uint32_t)) */,
243*4bdc9457SAndroid Build Coastguard Worker     &xnn_params.x32.zip);
244*4bdc9457SAndroid Build Coastguard Worker }
245