Skip to content

Commit a2e77e6

Browse files
airMengairMeng
committed
clang-format
1 parent 1bf2aa4 commit a2e77e6

File tree

9 files changed

+8888
-7046
lines changed

9 files changed

+8888
-7046
lines changed

‎ggml-sycl/backend.hpp‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -14,8 +14,8 @@
1414
#define GGML_SYCL_BACKEND_HPP
1515

1616
#include "common.hpp"
17-
#include "vecdotq.hpp"
18-
#include "mmvq.hpp"
1917
#include "mmq.hpp"
18+
#include "mmvq.hpp"
19+
#include "vecdotq.hpp"
2020

2121
#endif // GGML_SYCL_BACKEND_HPP

‎ggml-sycl/common.cpp‎

Lines changed: 161 additions & 139 deletions
Original file line numberDiff line numberDiff line change
@@ -12,169 +12,191 @@
1212

1313
#include "common.hpp"
1414

15-
int get_main_device(){
16-
return g_main_device;
15+
int get_main_device() {
16+
return g_main_device;
1717
}
1818

1919
void check_allow_gpu_index(const int device_index) {
20-
if (device_index >= g_device_count) {
21-
char error_buf[256];
22-
snprintf(error_buf, sizeof(error_buf),
23-
"%s error: device_index:%d is out of range: [0-%d]", __func__,
24-
device_index, g_device_count - 1);
25-
fprintf(stderr, "%s\n", error_buf);
26-
assert(false);
27-
}
20+
if (device_index >= g_device_count) {
21+
char error_buf[256];
22+
snprintf(
23+
error_buf,
24+
sizeof(error_buf),
25+
"%s error: device_index:%d is out of range: [0-%d]",
26+
__func__,
27+
device_index,
28+
g_device_count - 1);
29+
fprintf(stderr, "%s\n", error_buf);
30+
assert(false);
31+
}
2832
}
2933

3034
void check_allow_gpu_id(const int device_id) {
31-
if (!g_sycl_gpu_mgr->is_allowed_gpu(device_id)) {
32-
char error_buf[256];
33-
snprintf(error_buf, sizeof(error_buf),
34-
"error: cannot set device=%d, which is not allowed. Please "
35-
"set GPU ID in: [%s]",
36-
device_id, g_sycl_gpu_mgr->gpus_list.c_str());
37-
fprintf(stderr, "%s\n", error_buf);
38-
throw std::invalid_argument(error_buf);
39-
}
35+
if (!g_sycl_gpu_mgr->is_allowed_gpu(device_id)) {
36+
char error_buf[256];
37+
snprintf(
38+
error_buf,
39+
sizeof(error_buf),
40+
"error: cannot set device=%d, which is not allowed. Please "
41+
"set GPU ID in: [%s]",
42+
device_id,
43+
g_sycl_gpu_mgr->gpus_list.c_str());
44+
fprintf(stderr, "%s\n", error_buf);
45+
throw std::invalid_argument(error_buf);
46+
}
4047
}
4148

4249
int get_current_device_id() {
43-
return dpct::dev_mgr::instance().current_device_id();
50+
return dpct::dev_mgr::instance().current_device_id();
4451
}
4552

46-
void log_ggml_var_device(const char*name, float *src, size_t total_elements, bool src_on_device){
47-
if(!g_ggml_sycl_debug) return;
48-
if(!src){
49-
printf("GGML Tensor:%s skip to save for NULL pointer\n", name);
50-
return;
51-
}
52-
char filename[1024];
53-
sprintf(filename, "%s.txt", name);
54-
printf("GGML Tensor:%s save to %s\n", name, filename);
55-
56-
size_t total_size = total_elements*sizeof(float);
57-
float *local_buf = NULL;
58-
if(src_on_device) {
59-
local_buf = (float *) ggml_sycl_host_malloc(total_size);
60-
ggml_sycl_set_device(g_main_device);
61-
dpct::queue_ptr main_stream = g_syclStreams[g_main_device][0];
62-
main_stream->memcpy(local_buf, src, total_size).wait();
63-
}
64-
else {
65-
local_buf = (float *)src;
66-
}
67-
68-
std::ofstream logfile;
69-
logfile.open(filename);
70-
for(size_t i=0; i<total_elements; i++){
71-
logfile << local_buf[i] <<" ";
72-
if((i+1)%20 ==0) logfile <<std::endl;
73-
}
74-
logfile <<std::endl;
75-
logfile.close();
76-
77-
if(src_on_device) ggml_sycl_host_free(local_buf);
53+
void log_ggml_var_device(
54+
const char* name,
55+
float* src,
56+
size_t total_elements,
57+
bool src_on_device) {
58+
if (!g_ggml_sycl_debug)
59+
return;
60+
if (!src) {
61+
printf("GGML Tensor:%s skip to save for NULL pointer\n", name);
62+
return;
63+
}
64+
char filename[1024];
65+
sprintf(filename, "%s.txt", name);
66+
printf("GGML Tensor:%s save to %s\n", name, filename);
67+
68+
size_t total_size = total_elements * sizeof(float);
69+
float* local_buf = NULL;
70+
if (src_on_device) {
71+
local_buf = (float*)ggml_sycl_host_malloc(total_size);
72+
ggml_sycl_set_device(g_main_device);
73+
dpct::queue_ptr main_stream = g_syclStreams[g_main_device][0];
74+
main_stream->memcpy(local_buf, src, total_size).wait();
75+
} else {
76+
local_buf = (float*)src;
77+
}
78+
79+
std::ofstream logfile;
80+
logfile.open(filename);
81+
for (size_t i = 0; i < total_elements; i++) {
82+
logfile << local_buf[i] << " ";
83+
if ((i + 1) % 20 == 0)
84+
logfile << std::endl;
85+
}
86+
logfile << std::endl;
87+
logfile.close();
88+
89+
if (src_on_device)
90+
ggml_sycl_host_free(local_buf);
7891
}
7992

80-
void log_ggml_var_device_fp16(const char*name, sycl::half *src, size_t total_elements, bool src_on_device){
81-
if(!g_ggml_sycl_debug) return;
82-
if(!src){
83-
printf("GGML Tensor:%s skip to save for NULL pointer\n", name);
84-
return;
85-
}
86-
char filename[1024];
87-
sprintf(filename, "%s.txt", name);
88-
printf("GGML Tensor:%s save to %s\n", name, filename);
89-
90-
size_t total_size = total_elements*sizeof(sycl::half);
91-
sycl::half *local_buf = NULL;
92-
if(src_on_device) {
93-
local_buf = (sycl::half *) ggml_sycl_host_malloc(total_size);
94-
ggml_sycl_set_device(g_main_device);
95-
dpct::queue_ptr main_stream = g_syclStreams[g_main_device][0];
96-
main_stream->memcpy(local_buf, src, total_size).wait();
97-
}
98-
else {
99-
local_buf = (sycl::half *)src;
100-
}
101-
102-
std::ofstream logfile;
103-
logfile.open(filename);
104-
for(size_t i=0; i<total_elements; i++){
105-
logfile << local_buf[i] <<" ";
106-
if((i+1)%20 ==0) logfile <<std::endl;
107-
}
108-
logfile <<std::endl;
109-
logfile.close();
110-
111-
if(src_on_device) ggml_sycl_host_free(local_buf);
93+
void log_ggml_var_device_fp16(
94+
const char* name,
95+
sycl::half* src,
96+
size_t total_elements,
97+
bool src_on_device) {
98+
if (!g_ggml_sycl_debug)
99+
return;
100+
if (!src) {
101+
printf("GGML Tensor:%s skip to save for NULL pointer\n", name);
102+
return;
103+
}
104+
char filename[1024];
105+
sprintf(filename, "%s.txt", name);
106+
printf("GGML Tensor:%s save to %s\n", name, filename);
107+
108+
size_t total_size = total_elements * sizeof(sycl::half);
109+
sycl::half* local_buf = NULL;
110+
if (src_on_device) {
111+
local_buf = (sycl::half*)ggml_sycl_host_malloc(total_size);
112+
ggml_sycl_set_device(g_main_device);
113+
dpct::queue_ptr main_stream = g_syclStreams[g_main_device][0];
114+
main_stream->memcpy(local_buf, src, total_size).wait();
115+
} else {
116+
local_buf = (sycl::half*)src;
117+
}
118+
119+
std::ofstream logfile;
120+
logfile.open(filename);
121+
for (size_t i = 0; i < total_elements; i++) {
122+
logfile << local_buf[i] << " ";
123+
if ((i + 1) % 20 == 0)
124+
logfile << std::endl;
125+
}
126+
logfile << std::endl;
127+
logfile.close();
128+
129+
if (src_on_device)
130+
ggml_sycl_host_free(local_buf);
112131
}
113132

114-
void print_ggml_tensor(const char*name, struct ggml_tensor *src){
115-
if(!g_ggml_sycl_debug) return;
116-
if(!src){
117-
printf("GGML Tensor:%s skip to save for NULL pointer\n", name);
118-
return;
119-
}
120-
121-
size_t total_elements = ggml_nelements(src);
122-
123-
const bool src_on_device = src->backend == GGML_BACKEND_TYPE_GPU || src->backend == GGML_BACKEND_TYPE_GPU_SPLIT;
124-
float *src_data =NULL;
125-
if(src_on_device) {
126-
ggml_tensor_extra_gpu * src_extra = (ggml_tensor_extra_gpu *) src->extra;
127-
src_data = (float*)src_extra->data_device[g_main_device];
128-
}
129-
else {
130-
src_data = (float *)src->data;
131-
}
132-
133-
log_ggml_var_device(name, src_data, total_elements, src_on_device);
133+
void print_ggml_tensor(const char* name, struct ggml_tensor* src) {
134+
if (!g_ggml_sycl_debug)
135+
return;
136+
if (!src) {
137+
printf("GGML Tensor:%s skip to save for NULL pointer\n", name);
138+
return;
139+
}
140+
141+
size_t total_elements = ggml_nelements(src);
142+
143+
const bool src_on_device = src->backend == GGML_BACKEND_TYPE_GPU ||
144+
src->backend == GGML_BACKEND_TYPE_GPU_SPLIT;
145+
float* src_data = NULL;
146+
if (src_on_device) {
147+
ggml_tensor_extra_gpu* src_extra = (ggml_tensor_extra_gpu*)src->extra;
148+
src_data = (float*)src_extra->data_device[g_main_device];
149+
} else {
150+
src_data = (float*)src->data;
151+
}
152+
153+
log_ggml_var_device(name, src_data, total_elements, src_on_device);
134154
}
135155

136-
void log_tensor_with_cnt(const char* name, struct ggml_tensor * src, int stop_cnt) {
137-
stop_cnt = 4;
138-
if(log_file_name_idx>=stop_cnt) return;
139-
char filename[1280];
140-
sprintf(filename, "%s_%07d", name, log_file_name_idx);
141-
log_file_name_idx++;
142-
print_ggml_tensor(filename, src);
156+
void log_tensor_with_cnt(
157+
const char* name,
158+
struct ggml_tensor* src,
159+
int stop_cnt) {
160+
stop_cnt = 4;
161+
if (log_file_name_idx >= stop_cnt)
162+
return;
163+
char filename[1280];
164+
sprintf(filename, "%s_%07d", name, log_file_name_idx);
165+
log_file_name_idx++;
166+
print_ggml_tensor(filename, src);
143167
}
144168

145-
void *ggml_sycl_host_malloc(size_t size) try {
146-
if (getenv("GGML_SYCL_NO_PINNED") != nullptr) {
147-
return nullptr;
148-
}
149-
150-
void * ptr = nullptr;
151-
//allow to use dpct::get_in_order_queue() for host malloc
152-
dpct::err0 err = CHECK_TRY_ERROR(
153-
ptr = (void *)sycl::malloc_host(size, dpct::get_in_order_queue()));
154-
155-
if (err != 0) {
156-
// clear the error
157-
fprintf(
158-
stderr,
159-
"WARNING: failed to allocate %.2f MB of pinned memory: %s\n",
160-
size / 1024.0 / 1024.0,
161-
"syclGetErrorString is not supported");
162-
return nullptr;
163-
}
164-
165-
return ptr;
166-
}
167-
catch (sycl::exception const &exc) {
169+
void* ggml_sycl_host_malloc(size_t size) try {
170+
if (getenv("GGML_SYCL_NO_PINNED") != nullptr) {
171+
return nullptr;
172+
}
173+
174+
void* ptr = nullptr;
175+
// allow to use dpct::get_in_order_queue() for host malloc
176+
dpct::err0 err = CHECK_TRY_ERROR(
177+
ptr = (void*)sycl::malloc_host(size, dpct::get_in_order_queue()));
178+
179+
if (err != 0) {
180+
// clear the error
181+
fprintf(
182+
stderr,
183+
"WARNING: failed to allocate %.2f MB of pinned memory: %s\n",
184+
size / 1024.0 / 1024.0,
185+
"syclGetErrorString is not supported");
186+
return nullptr;
187+
}
188+
189+
return ptr;
190+
} catch (sycl::exception const& exc) {
168191
std::cerr << exc.what() << "Exception caught at file:" << __FILE__
169192
<< ", line:" << __LINE__ << std::endl;
170193
std::exit(1);
171194
}
172195

173-
void ggml_sycl_host_free(void *ptr) try {
174-
//allow to use dpct::get_in_order_queue() for host malloc
175-
SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(ptr, dpct::get_in_order_queue())));
176-
}
177-
catch (sycl::exception const &exc) {
196+
void ggml_sycl_host_free(void* ptr) try {
197+
// allow to use dpct::get_in_order_queue() for host malloc
198+
SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(ptr, dpct::get_in_order_queue())));
199+
} catch (sycl::exception const& exc) {
178200
std::cerr << exc.what() << "Exception caught at file:" << __FILE__
179201
<< ", line:" << __LINE__ << std::endl;
180202
std::exit(1);

0 commit comments

Comments
 (0)