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
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
|
#include "ggml.h"
#include "ggml-cpu.h"
#include "ggml-alloc.h"
#include "ggml-backend.h"
#ifdef GGML_USE_CUDA
#include "ggml-cuda.h"
#endif
#ifdef GGML_USE_METAL
#include "ggml-metal.h"
#endif
#include <cassert>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <map>
#include <string>
#include <vector>
static void ggml_log_callback_default(ggml_log_level level, const char * text, void * user_data) {
(void) level;
(void) user_data;
fputs(text, stderr);
fflush(stderr);
}
struct test_model {
struct ggml_tensor * a;
struct ggml_tensor * b;
ggml_backend_t backend = NULL;
ggml_backend_buffer_t buffer;
struct ggml_context * ctx;
};
void load_model(test_model & model, bool use_gpu = false) {
// create data
int K = 3, IC = 10, OC = 10;
int IL = 8, N = 1;
// Initialize adata
std::vector<float> adata(K * IC * OC);
for (int i = 0; i < K * IC * OC; i++) {
adata[i] = 4.5f;
}
// Convert adata to fp16 format
std::vector<ggml_fp16_t> hadata(K * IC * OC);
ggml_fp32_to_fp16_row(adata.data(), hadata.data(), K * IC * OC);
// Initialize bdata
std::vector<float> bdata(IL * IC * N);
for (int i = 0; i < IL * IC * N; i++) {
bdata[i] = 2.5f;
}
size_t buffer_size = 0;
{
buffer_size += K * IC * OC * ggml_type_size(GGML_TYPE_F16); // tensor a
buffer_size += IL * IC * N * ggml_type_size(GGML_TYPE_F32); // tensor b
buffer_size += 1024; // overhead
}
printf("%s: ggml tensor size = %d bytes\n", __func__, (int) sizeof(ggml_tensor));
printf("%s: backend buffer size = %0.2f MB\n", __func__, (buffer_size/ 1024.f/ 1024.f));
ggml_log_set(ggml_log_callback_default, nullptr);
int num_tensors = 2;
struct ggml_init_params params {
/*.mem_size =*/ ggml_tensor_overhead() * num_tensors,
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true,
};
// initialize the backend
#ifdef GGML_USE_CUDA
if (use_gpu) {
fprintf(stderr, "%s: using CUDA backend\n", __func__);
model.backend = ggml_backend_cuda_init(0);
if (!model.backend) {
fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__);
}
}
#endif
#ifdef GGML_USE_METAL
if (use_gpu) {
fprintf(stderr, "%s: using Metal backend\n", __func__);
model.backend = ggml_backend_metal_init();
if (!model.backend) {
fprintf(stderr, "%s: ggml_backend_metal_init() failed\n", __func__);
}
}
#endif
if(!model.backend) {
// fallback to CPU backend
model.backend = ggml_backend_cpu_init();
}
model.buffer = ggml_backend_alloc_buffer(model.backend, buffer_size);
// create context
model.ctx = ggml_init(params);
// create tensors
model.a = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F16, K, IC, OC);
model.b = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, IL, IC, N);
// create a allocator
ggml_tallocr alloc = ggml_tallocr_new(model.buffer);
// alloc memory
ggml_tallocr_alloc(&alloc, model.a);
// load data to buffer
if(ggml_backend_is_cpu(model.backend)) {
memcpy(model.a->data, hadata.data(), ggml_nbytes(model.a));
} else {
ggml_backend_tensor_set(model.a, hadata.data(), 0, ggml_nbytes(model.a));
}
// alloc memory
ggml_tallocr_alloc(&alloc, model.b);
if(ggml_backend_is_cpu(model.backend)
#ifdef GGML_USE_METAL
|| ggml_backend_is_metal(model.backend)
#endif
) {
memcpy(model.b->data, bdata.data(), ggml_nbytes(model.b));
} else {
ggml_backend_tensor_set(model.b, bdata.data(), 0, ggml_nbytes(model.b));
}
}
struct ggml_cgraph * build_graph(const test_model& model) {
static size_t buf_size = ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead();
static std::vector<uint8_t> buf(buf_size);
struct ggml_init_params params0 = {
/*.mem_size =*/ buf_size,
/*.mem_buffer =*/ buf.data(),
/*.no_alloc =*/ true, // the tensors will be allocated later by ggml_gallocr_alloc_graph()
};
// create a temporally context to build the graph
struct ggml_context * ctx0 = ggml_init(params0);
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
int s0 = 1;
int p0 = 1;
int d0 = 1;
// split conv1d in fundamental methods for test unit
struct ggml_tensor* im2col_0 = ggml_im2col(ctx0, model.a, model.b, s0, 0, p0, 0, d0, 0, false, GGML_TYPE_F16);
ggml_set_name(im2col_0, "im2col_res");
ggml_build_forward_expand(gf, im2col_0);
struct ggml_tensor* conv1d_res = ggml_conv_1d(ctx0, model.a, model.b, s0, p0, d0);
ggml_set_name(conv1d_res, "conv1d_res");
ggml_build_forward_expand(gf, conv1d_res);
// delete the temporally context used to build the graph
ggml_free(ctx0);
return gf;
}
struct ggml_cgraph* compute_graph(const test_model & model, ggml_gallocr_t allocr) {
struct ggml_cgraph * gf = build_graph(model);
// allocate tensors
ggml_gallocr_alloc_graph(allocr, gf);
int n_threads = 1;
if (ggml_backend_is_cpu(model.backend)) {
ggml_backend_cpu_set_n_threads(model.backend, n_threads);
}
ggml_backend_graph_compute(model.backend, gf);
//ggml_graph_print(gf);
return gf;
}
int main(void)
{
ggml_time_init();
test_model model;
load_model(model, true);
ggml_gallocr_t allocr = NULL;
{
allocr = ggml_gallocr_new(ggml_backend_get_default_buffer_type(model.backend));
//create the worst case graph for memory usage estimation
struct ggml_cgraph * gf = build_graph(model);
// compute the required memory
ggml_gallocr_reserve(allocr, gf);
size_t mem_size = ggml_gallocr_get_buffer_size(allocr, 0);
fprintf(stderr, "%s: compute buffer size: %.2f MB\n", __func__, mem_size/1024.0f/1024.0f);
}
struct ggml_cgraph * gf_res = compute_graph(model, allocr);
struct ggml_tensor * im2col_res = NULL;
struct ggml_tensor * conv1d_res = NULL;
for(int i = 0; i < ggml_graph_n_nodes(gf_res); i++) {
if(strcmp(ggml_get_name(ggml_graph_node(gf_res, i)), "im2col_res") == 0) {
im2col_res = ggml_graph_node(gf_res, i);
} else if(strcmp(ggml_get_name(ggml_graph_node(gf_res, i)), "conv1d_res") == 0) {
conv1d_res = ggml_graph_node(gf_res, i);
}
}
std::vector<uint16_t> im2col_data(ggml_nelements(im2col_res));
std::vector<float> conv2d_data(ggml_nelements(conv1d_res));
ggml_backend_tensor_get(im2col_res, im2col_data.data(), 0, ggml_nbytes(im2col_res));
ggml_backend_tensor_get(conv1d_res, conv2d_data.data(), 0, ggml_nbytes(conv1d_res));
const int n_conv1d_test = 80;
const int n_im2col_test = 240;
float expected_conv1d[n_conv1d_test] = {
225.00f, 337.50f, 337.50f, 337.50f, 337.50f, 337.50f, 337.50f, 225.00f,
225.00f, 337.50f, 337.50f, 337.50f, 337.50f, 337.50f, 337.50f, 225.00f,
225.00f, 337.50f, 337.50f, 337.50f, 337.50f, 337.50f, 337.50f, 225.00f,
225.00f, 337.50f, 337.50f, 337.50f, 337.50f, 337.50f, 337.50f, 225.00f,
225.00f, 337.50f, 337.50f, 337.50f, 337.50f, 337.50f, 337.50f, 225.00f,
225.00f, 337.50f, 337.50f, 337.50f, 337.50f, 337.50f, 337.50f, 225.00f,
225.00f, 337.50f, 337.50f, 337.50f, 337.50f, 337.50f, 337.50f, 225.00f,
225.00f, 337.50f, 337.50f, 337.50f, 337.50f, 337.50f, 337.50f, 225.00f,
225.00f, 337.50f, 337.50f, 337.50f, 337.50f, 337.50f, 337.50f, 225.00f,
225.00f, 337.50f, 337.50f, 337.50f, 337.50f, 337.50f, 337.50f, 225.00f
};
// first im2col test
uint16_t expected_im2col[n_conv1d_test] = {
0, 16640, 16640, 0, 16640, 16640, 0, 16640,
16640, 0, 16640, 16640, 0, 16640, 16640, 0,
16640, 16640, 0, 16640, 16640, 0, 16640, 16640,
0, 16640, 16640, 0, 16640, 16640, 16640, 16640,
16640, 16640, 16640, 16640, 16640, 16640, 16640, 16640,
16640, 16640, 16640, 16640, 16640, 16640, 16640, 16640,
16640, 16640, 16640, 16640, 16640, 16640, 16640, 16640,
16640, 16640, 16640, 16640, 16640, 16640, 16640, 16640,
16640, 16640, 16640, 16640, 16640, 16640, 16640, 16640,
16640, 16640, 16640, 16640, 16640, 16640, 16640, 16640
};
printf("\nPerforming test:\n");
bool passed = true;
for(int i = 0; i < n_conv1d_test; i++) {
if(
im2col_data[i] != expected_im2col[i]) {
passed = false;
break;
}
}
printf("ggml_im2col (%d): %s\n", (int) ggml_nelements(im2col_res), passed && (ggml_nelements(im2col_res) == n_im2col_test) ? "\033[32mPASSED\033[0m" : "\033[31mFAILED\033[0m");
passed = true;
for(int i = 0; i < n_conv1d_test; i++) {
if(conv2d_data[i] != expected_conv1d[i]) {
passed = false;
break;
}
}
printf("ggml_conv1d (%d): %s\n", (int) ggml_nelements(conv1d_res), passed && (ggml_nelements(conv1d_res) == n_conv1d_test) ? "\033[32mPASSED\033[0m" : "\033[31mFAILED\033[0m");
ggml_free(model.ctx);
ggml_backend_buffer_free(model.buffer);
ggml_backend_free(model.backend);
ggml_gallocr_free(allocr);
return 0;
}
|
