1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
|
#pragma clang diagnostic ignored "-Wunused-variable"
#pragma clang diagnostic ignored "-Wunused-function"
#pragma clang diagnostic ignored "-Wunused-but-set-variable"
#include <HAP_farf.h>
#include <HAP_perf.h>
#include <math.h>
#include <string.h>
#define GGML_COMMON_DECL_C
#include "ggml-common.h"
#include "htp-ctx.h"
#include "htp-msg.h"
#include "htp-ops.h"
#include "hvx-utils.h"
struct htp_copy_context {
struct htp_ops_context * octx;
uint32_t src0_type_size;
uint32_t src0_block_size;
uint32_t dst_type_size;
uint32_t dst_block_size;
uint32_t src0_blocks_per_row;
uint32_t dst_blocks_per_row;
uint32_t src0_nrows_per_thread;
void (*copy)(struct htp_copy_context * ct, struct htp_ops_context * octx, int nth, int ith);
};
#define cpy_preamble \
struct htp_tensor *src0 = &octx->src0; \
struct htp_tensor *dst = &octx->dst; \
\
const uint32_t ne00 = src0->ne[0]; \
const uint32_t ne01 = src0->ne[1]; \
const uint32_t ne02 = src0->ne[2]; \
const uint32_t ne03 = src0->ne[3]; \
\
const uint32_t nb00 = src0->nb[0]; \
const uint32_t nb01 = src0->nb[1]; \
const uint32_t nb02 = src0->nb[2]; \
const uint32_t nb03 = src0->nb[3]; \
\
const uint32_t ne0 = dst->ne[0]; \
const uint32_t ne1 = dst->ne[1]; \
const uint32_t ne2 = dst->ne[2]; \
const uint32_t ne3 = dst->ne[3]; \
\
const uint32_t nb0 = dst->nb[0]; \
const uint32_t nb1 = dst->nb[1]; \
const uint32_t nb2 = dst->nb[2]; \
const uint32_t nb3 = dst->nb[3]; \
\
const uint32_t nr = ne01;
static void cpy_thread_sametype_sameshape(struct htp_copy_context * ct, struct htp_ops_context * octx, const int nth, const int ith) {
cpy_preamble;
// parallelize by src0 rows
const uint32_t dr = ct->src0_nrows_per_thread;
const uint32_t ir0 = dr * ith;
const uint32_t ir1 = (ir0 + dr) < nr ? (ir0 + dr) : nr;
// copy by rows
for (uint32_t i03 = 0; i03 < ne03; i03++) {
for (uint32_t i02 = 0; i02 < ne02; i02++) {
#pragma unroll(2)
for (uint32_t i01 = ir0; i01 < ir1; i01++) {
uint8_t* dst_ptr = (uint8_t*) dst->data + i01*nb1 + i02*nb2 + i03*nb3;
uint8_t* src0_ptr = (uint8_t*) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
hex_l2fetch(src0_ptr, ne00 * ct->src0_type_size, nb01, 2);
hvx_copy_uu(dst_ptr, src0_ptr, ne00, ct->src0_type_size);
}
}
}
}
static void cpy_thread_sametype_reshape(struct htp_copy_context * ct, struct htp_ops_context * octx, int nth, int ith) {
cpy_preamble;
// parallelize by src0 rows
const uint32_t dr = ct->src0_nrows_per_thread;
const uint32_t ir0 = dr * ith;
const uint32_t ir1 = (ir0 + dr) < nr ? (ir0 + dr) : nr;
// dst counters
int64_t k10 = 0;
int64_t i11 = 0;
int64_t i12 = 0;
int64_t i13 = 0;
// number of blocks in a row
const int64_t nk00 = ct->src0_blocks_per_row;
const int64_t nk0 = ct->dst_blocks_per_row;
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
k10 += nk00 * ir0;
while (k10 >= nk0) {
k10 -= nk0;
if (++i11 == ne1) {
i11 = 0;
if (++i12 == ne2) {
i12 = 0;
if (++i13 == ne3) {
i13 = 0;
}
}
}
}
for (int64_t i01 = ir0; i01 < ir1; i01++) {
for (int64_t k00 = 0; k00 < nk00; k00++) {
const char * src0_ptr = ((char *) src0->data + k00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
char * dst_ptr = ((char *) dst->data + k10*nb0 + i11*nb1 + i12*nb2 + i13*nb3);
memcpy(dst_ptr, src0_ptr, ct->dst_type_size);
if (++k10 == nk0) {
k10 = 0;
if (++i11 == ne1) {
i11 = 0;
if (++i12 == ne2) {
i12 = 0;
if (++i13 == ne3) {
i13 = 0;
}
}
}
}
}
}
k10 += nk00 * (ne01 - ir1);
while (k10 >= nk0) {
k10 -= nk0;
if (++i11 == ne1) {
i11 = 0;
if (++i12 == ne2) {
i12 = 0;
if (++i13 == ne3) {
i13 = 0;
}
}
}
}
}
}
}
static void cpy_thread_f16_f32_sameshape(struct htp_copy_context * ct, struct htp_ops_context * octx, const int nth, const int ith) {
cpy_preamble;
// parallelize by src0 rows
const uint32_t dr = ct->src0_nrows_per_thread;
const uint32_t ir0 = dr * ith;
const uint32_t ir1 = (ir0 + dr) < nr ? (ir0 + dr) : nr;
// copy by rows
for (uint32_t i03 = 0; i03 < ne03; i03++) {
for (uint32_t i02 = 0; i02 < ne02; i02++) {
#pragma unroll(2)
for (uint32_t i01 = ir0; i01 < ir1; i01++) {
uint8_t* dst_ptr = (uint8_t*) dst->data + i01*nb1 + i02*nb2 + i03*nb3;
uint8_t* src0_ptr = (uint8_t*) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
hex_l2fetch(src0_ptr, ne00 * sizeof(float), nb01, 2);
hvx_copy_f16_f32_uu(dst_ptr, src0_ptr, ne00);
}
}
}
}
static void cpy_thread_f32_f16_sameshape(struct htp_copy_context * ct, struct htp_ops_context * octx, const int nth, const int ith) {
cpy_preamble;
// parallelize by src0 rows
const uint32_t dr = ct->src0_nrows_per_thread;
const uint32_t ir0 = dr * ith;
const uint32_t ir1 = (ir0 + dr) < nr ? (ir0 + dr) : nr;
// copy by rows
for (uint32_t i03 = 0; i03 < ne03; i03++) {
for (uint32_t i02 = 0; i02 < ne02; i02++) {
#pragma unroll(2)
for (uint32_t i01 = ir0; i01 < ir1; i01++) {
uint8_t* dst_ptr = (uint8_t*) dst->data + i01*nb1 + i02*nb2 + i03*nb3;
uint8_t* src0_ptr = (uint8_t*) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
hex_l2fetch(src0_ptr, ne00 * sizeof(__fp16), nb01, 2);
hvx_copy_f32_f16_uu(dst_ptr, src0_ptr, ne00);
}
}
}
}
static void cpy_work_func(unsigned int n, unsigned int i, void *data) {
struct htp_copy_context *ct = (struct htp_copy_context *) data;
ct->copy(ct, ct->octx, n, i);
}
int op_cpy(struct htp_ops_context * octx) {
cpy_preamble;
struct htp_copy_context ct;
ct.octx = octx;
switch (src0->type) {
case HTP_TYPE_F32: ct.src0_type_size = 4; ct.src0_block_size = 1; ct.src0_blocks_per_row = ne00 / 1; break;
case HTP_TYPE_F16: ct.src0_type_size = 2; ct.src0_block_size = 1; ct.src0_blocks_per_row = ne00 / 1; break;
default:
return HTP_STATUS_NO_SUPPORT;
}
switch (dst->type) {
case HTP_TYPE_F32: ct.dst_type_size = 4; ct.dst_block_size = 1; ct.dst_blocks_per_row = ne0 / 1; break;
case HTP_TYPE_F16: ct.dst_type_size = 2; ct.dst_block_size = 1; ct.dst_blocks_per_row = ne0 / 1; break;
default:
return HTP_STATUS_NO_SUPPORT;
}
if (octx->flags & HTP_OPFLAGS_SKIP_COMPUTE) {
return HTP_STATUS_OK;
}
const bool sametype = (src0->type == dst->type);
const bool transposed = (nb00 > nb01) || (nb0 > nb1);
const bool sameshape = !transposed && (ne00 == ne0 && ne01 == ne1 && ne02 == ne2 && ne03 == ne3);
const uint32_t n_jobs = MIN(nr, octx->n_threads);
ct.src0_nrows_per_thread = (nr + n_jobs - 1) / n_jobs;
if (sametype && sameshape) {
ct.copy = cpy_thread_sametype_sameshape;
} else if (sameshape) {
/**/ if (dst->type == HTP_TYPE_F16 && src0->type == HTP_TYPE_F32)
ct.copy = cpy_thread_f16_f32_sameshape;
else if (dst->type == HTP_TYPE_F32 && src0->type == HTP_TYPE_F16)
ct.copy = cpy_thread_f32_f16_sameshape;
else
return HTP_STATUS_NO_SUPPORT;
} else if (sametype) {
ct.copy = cpy_thread_sametype_reshape;
} else {
return HTP_STATUS_NO_SUPPORT;
}
worker_pool_run_func(octx->ctx->worker_pool, cpy_work_func, &ct, n_jobs);
return HTP_STATUS_OK;
}
|
