|
10879
|
1 /*
|
|
|
2 * Copyright (c) 2013, Oracle and/or its affiliates. All rights reserved.
|
|
|
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
|
|
|
4 *
|
|
|
5 * This code is free software; you can redistribute it and/or modify it
|
|
|
6 * under the terms of the GNU General Public License version 2 only, as
|
|
|
7 * published by the Free Software Foundation.
|
|
|
8 *
|
|
|
9 * This code is distributed in the hope that it will be useful, but WITHOUT
|
|
|
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
|
|
|
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
|
|
|
12 * version 2 for more details (a copy is included in the LICENSE file that
|
|
|
13 * accompanied this code).
|
|
|
14 *
|
|
|
15 * You should have received a copy of the GNU General Public License version
|
|
|
16 * 2 along with this work; if not, write to the Free Software Foundation,
|
|
|
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
|
|
|
18 *
|
|
|
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
|
|
|
20 * or visit www.oracle.com if you need additional information or have any
|
|
|
21 * questions.
|
|
|
22 *
|
|
|
23 */
|
|
|
24
|
|
|
25 #include "precompiled.hpp"
|
|
|
26 #include "runtime/gpu.hpp"
|
|
|
27 #include "utilities/globalDefinitions.hpp"
|
|
|
28 #include "utilities/ostream.hpp"
|
|
|
29
|
|
|
30 void * gpu::Ptx::_device_context;
|
|
|
31
|
|
|
32 gpu::Ptx::cuda_cu_init_func_t gpu::Ptx::_cuda_cu_init;
|
|
|
33 gpu::Ptx::cuda_cu_ctx_create_func_t gpu::Ptx::_cuda_cu_ctx_create;
|
|
|
34 gpu::Ptx::cuda_cu_ctx_detach_func_t gpu::Ptx::_cuda_cu_ctx_detach;
|
|
|
35 gpu::Ptx::cuda_cu_ctx_synchronize_func_t gpu::Ptx::_cuda_cu_ctx_synchronize;
|
|
|
36 gpu::Ptx::cuda_cu_device_get_count_func_t gpu::Ptx::_cuda_cu_device_get_count;
|
|
|
37 gpu::Ptx::cuda_cu_device_get_name_func_t gpu::Ptx::_cuda_cu_device_get_name;
|
|
|
38 gpu::Ptx::cuda_cu_device_get_func_t gpu::Ptx::_cuda_cu_device_get;
|
|
|
39 gpu::Ptx::cuda_cu_device_compute_capability_func_t gpu::Ptx::_cuda_cu_device_compute_capability;
|
|
|
40 gpu::Ptx::cuda_cu_device_get_attribute_func_t gpu::Ptx::_cuda_cu_device_get_attribute;
|
|
|
41 gpu::Ptx::cuda_cu_launch_kernel_func_t gpu::Ptx::_cuda_cu_launch_kernel;
|
|
|
42 gpu::Ptx::cuda_cu_module_get_function_func_t gpu::Ptx::_cuda_cu_module_get_function;
|
|
|
43 gpu::Ptx::cuda_cu_module_load_data_ex_func_t gpu::Ptx::_cuda_cu_module_load_data_ex;
|
|
|
44
|
|
|
45 void gpu::probe_linkage() {
|
|
|
46 #if defined(__APPLE__) || defined(LINUX)
|
|
|
47 set_gpu_linkage(gpu::Ptx::probe_linkage());
|
|
|
48 #else
|
|
|
49 set_gpu_linkage(false);
|
|
|
50 #endif
|
|
|
51 }
|
|
|
52
|
|
|
53 void gpu::initialize_gpu() {
|
|
|
54 if (gpu::has_gpu_linkage()) {
|
|
|
55 set_initialized(gpu::Ptx::initialize_gpu());
|
|
|
56 }
|
|
|
57 }
|
|
|
58
|
|
|
59 void * gpu::generate_kernel(unsigned char *code, int code_len, const char *name) {
|
|
|
60 if (gpu::has_gpu_linkage()) {
|
|
|
61 return (gpu::Ptx::generate_kernel(code, code_len, name));
|
|
|
62 } else {
|
|
|
63 return NULL;
|
|
|
64 }
|
|
|
65 }
|
|
|
66
|
|
|
67 bool gpu::execute_kernel(address kernel) {
|
|
|
68 if (gpu::has_gpu_linkage()) {
|
|
|
69 return (gpu::Ptx::execute_kernel(kernel));
|
|
|
70 } else {
|
|
|
71 return false;
|
|
|
72 }
|
|
|
73 }
|
|
|
74
|
|
|
75 bool gpu::Ptx::initialize_gpu() {
|
|
|
76
|
|
|
77 /* Initialize CUDA driver API */
|
|
|
78 int status = _cuda_cu_init(0);
|
|
|
79 if (status != GRAAL_CUDA_SUCCESS) {
|
|
|
80 tty->print_cr("Failed to initialize CUDA device");
|
|
|
81 return false;
|
|
|
82 }
|
|
|
83
|
|
|
84 if (TraceGPUInteraction) {
|
|
|
85 tty->print_cr("CUDA driver initialization: Success");
|
|
|
86 }
|
|
|
87
|
|
|
88 /* Get the number of compute-capable device count */
|
|
|
89 int device_count = 0;
|
|
|
90 status = _cuda_cu_device_get_count(&device_count);
|
|
|
91 if (status != GRAAL_CUDA_SUCCESS) {
|
|
|
92 tty->print_cr("[CUDA] Failed to get compute-capable device count");
|
|
|
93 return false;
|
|
|
94 }
|
|
|
95
|
|
|
96 if (device_count == 0) {
|
|
|
97 tty->print_cr("[CUDA] Found no device supporting CUDA");
|
|
|
98 return false;
|
|
|
99 }
|
|
|
100
|
|
|
101 if (TraceGPUInteraction) {
|
|
|
102 tty->print_cr("[CUDA] Number of compute-capable devices found: %d", device_count);
|
|
|
103 }
|
|
|
104
|
|
|
105 /* Get the handle to the first compute device */
|
|
|
106 int device_id = 0;
|
|
|
107 /* Compute-capable device handle */
|
|
|
108 int cu_device = 0;
|
|
|
109 status = _cuda_cu_device_get(&cu_device, device_id);
|
|
|
110
|
|
|
111 if (status != GRAAL_CUDA_SUCCESS) {
|
|
|
112 tty->print_cr("[CUDA] Failed to get handle of first compute-capable device i.e., the one at ordinal: %d", device_id);
|
|
|
113 return false;
|
|
|
114 }
|
|
|
115
|
|
|
116 if (TraceGPUInteraction) {
|
|
|
117 tty->print_cr("[CUDA] Got the handle of first compute-device");
|
|
|
118 }
|
|
|
119
|
|
|
120 /* Get device attributes */
|
|
|
121 int minor, major;
|
|
|
122 status = _cuda_cu_device_get_attribute(&minor, GRAAL_CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, cu_device);
|
|
|
123
|
|
|
124 if (status != GRAAL_CUDA_SUCCESS) {
|
|
|
125 tty->print_cr("[CUDA] Failed to get minor attribute of device: %d", cu_device);
|
|
|
126 return false;
|
|
|
127 }
|
|
|
128
|
|
|
129 status = _cuda_cu_device_get_attribute(&major, GRAAL_CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, cu_device);
|
|
|
130
|
|
|
131 if (status != GRAAL_CUDA_SUCCESS) {
|
|
|
132 tty->print_cr("[CUDA] Failed to get major attribute of device: %d", cu_device);
|
|
|
133 return false;
|
|
|
134 }
|
|
|
135
|
|
|
136 if (TraceGPUInteraction) {
|
|
|
137 tty->print_cr("[CUDA] Compatibility version of device %d: %d.%d", cu_device, major, minor);
|
|
|
138 }
|
|
|
139
|
|
|
140 /* Get device name */
|
|
|
141 char device_name[256];
|
|
|
142 status = _cuda_cu_device_get_name(device_name, 256, cu_device);
|
|
|
143
|
|
|
144 if (status != GRAAL_CUDA_SUCCESS) {
|
|
|
145 tty->print_cr("[CUDA] Failed to get name of device: %d", cu_device);
|
|
|
146 return false;
|
|
|
147 }
|
|
|
148
|
|
|
149 if (TraceGPUInteraction) {
|
|
|
150 tty->print_cr("[CUDA] Using %s", device_name);
|
|
|
151 }
|
|
|
152
|
|
|
153 /* Create CUDA context */
|
|
|
154 status = _cuda_cu_ctx_create(&_device_context, 0, cu_device);
|
|
|
155
|
|
|
156 if (status != GRAAL_CUDA_SUCCESS) {
|
|
|
157 tty->print_cr("[CUDA] Failed to create CUDA context for device: %d", cu_device);
|
|
|
158 return false;
|
|
|
159 }
|
|
|
160
|
|
|
161 if (TraceGPUInteraction) {
|
|
|
162 tty->print_cr("[CUDA] Success: Created context for device: %d", cu_device);
|
|
|
163 }
|
|
|
164
|
|
|
165 return true;
|
|
|
166 }
|
|
|
167
|
|
|
168 void *gpu::Ptx::generate_kernel(unsigned char *code, int code_len, const char *name) {
|
|
|
169
|
|
|
170 struct CUmod_st * cu_module;
|
|
|
171 // Use three JIT compiler options
|
|
|
172 const unsigned int jit_num_options = 3;
|
|
|
173 int *jit_options = NEW_C_HEAP_ARRAY(int, jit_num_options, mtCompiler);
|
|
|
174 void **jit_option_values = NEW_C_HEAP_ARRAY(void *, jit_num_options, mtCompiler);
|
|
|
175
|
|
|
176 // Set up PTX JIT compiler options
|
|
|
177 // 1. set size of compilation log buffer
|
|
|
178 int jit_log_buffer_size = 1024;
|
|
|
179 jit_options[0] = GRAAL_CU_JIT_INFO_LOG_BUFFER_SIZE_BYTES;
|
|
|
180 jit_option_values[0] = (void *)(size_t)jit_log_buffer_size;
|
|
|
181
|
|
|
182 // 2. set pointer to compilation log buffer
|
|
|
183 char *jit_log_buffer = NEW_C_HEAP_ARRAY(char, jit_log_buffer_size, mtCompiler);
|
|
|
184 jit_options[1] = GRAAL_CU_JIT_INFO_LOG_BUFFER;
|
|
|
185 jit_option_values[1] = jit_log_buffer;
|
|
|
186
|
|
|
187 // 3. set pointer to set the Maximum # of registers (32) for the kernel
|
|
|
188 int jit_register_count = 32;
|
|
|
189 jit_options[2] = GRAAL_CU_JIT_MAX_REGISTERS;
|
|
|
190 jit_option_values[2] = (void *)(size_t)jit_register_count;
|
|
|
191
|
|
|
192 if (TraceGPUInteraction) {
|
|
|
193 tty->print_cr("[CUDA] PTX Kernel\n%s", code);
|
|
|
194 tty->print_cr("[CUDA] Function name : %s", name);
|
|
|
195
|
|
|
196 }
|
|
|
197
|
|
|
198 /* Load module's data with compiler options */
|
|
|
199 int status = _cuda_cu_module_load_data_ex(&cu_module, code, jit_num_options,
|
|
|
200 jit_options, (void **)jit_option_values);
|
|
|
201 if (status != GRAAL_CUDA_SUCCESS) {
|
|
|
202 if (status == GRAAL_CUDA_ERROR_NO_BINARY_FOR_GPU) {
|
|
|
203 tty->print_cr("[CUDA] Check for malformed PTX kernel or incorrect PTX compilation options");
|
|
|
204 }
|
|
|
205 tty->print_cr("[CUDA] *** Error (%d) Failed to load module data with online compiler options for method %s",
|
|
|
206 status, name);
|
|
|
207 return NULL;
|
|
|
208 }
|
|
|
209
|
|
|
210 if (TraceGPUInteraction) {
|
|
|
211 tty->print_cr("[CUDA] Loaded data for PTX Kernel");
|
|
|
212 }
|
|
|
213
|
|
|
214 struct CUfunc_st * cu_function;
|
|
|
215
|
|
|
216 status = _cuda_cu_module_get_function(&cu_function, cu_module, name);
|
|
|
217
|
|
|
218 if (status != GRAAL_CUDA_SUCCESS) {
|
|
|
219 tty->print_cr("[CUDA] Failed to get function %s", name);
|
|
|
220 return NULL;
|
|
|
221 }
|
|
|
222
|
|
|
223 if (TraceGPUInteraction) {
|
|
|
224 tty->print_cr("[CUDA] Got function handle for %s", name);
|
|
|
225 }
|
|
|
226 return cu_function;
|
|
|
227 }
|
|
|
228
|
|
|
229 bool gpu::Ptx::execute_kernel(address kernel) {
|
|
|
230 // grid dimensionality
|
|
|
231 unsigned int gridX = 1;
|
|
|
232 unsigned int gridY = 1;
|
|
|
233 unsigned int gridZ = 1;
|
|
|
234
|
|
|
235 // thread dimensionality
|
|
|
236 unsigned int blockX = 1;
|
|
|
237 unsigned int blockY = 1;
|
|
|
238 unsigned int blockZ = 1;
|
|
|
239
|
|
|
240 int *cu_function = (int *)kernel;
|
|
|
241
|
|
|
242 if (kernel == NULL) {
|
|
|
243 return false;
|
|
|
244 }
|
|
|
245
|
|
|
246 if (TraceGPUInteraction) {
|
|
|
247 tty->print_cr("[CUDA] launching kernel");
|
|
|
248 }
|
|
|
249 int status = _cuda_cu_launch_kernel(cu_function,
|
|
|
250 gridX, gridY, gridZ,
|
|
|
251 blockX, blockY, blockZ,
|
|
|
252 0, NULL, NULL, NULL);
|
|
|
253 if (status != GRAAL_CUDA_SUCCESS) {
|
|
|
254 tty->print_cr("[CUDA] Failed to launch kernel");
|
|
|
255 return false;
|
|
|
256 }
|
|
|
257
|
|
|
258 if (TraceGPUInteraction) {
|
|
|
259 tty->print_cr("[CUDA] Success: Kernel Launch");
|
|
|
260 }
|
|
|
261 return status == 0; // GRAAL_CUDA_SUCCESS
|
|
|
262 }
|
|
|
263
|
|
|
264 #if defined(LINUX)
|
|
|
265 static const char cuda_library_name[] = "libcuda.so";
|
|
|
266 #elif defined(__APPLE__)
|
|
|
267 static char const cuda_library_name[] = "/usr/local/cuda/lib/libcuda.dylib";
|
|
|
268 #else
|
|
|
269 static char const cuda_library_name[] = "";
|
|
|
270 #endif
|
|
|
271
|
|
|
272 bool gpu::Ptx::probe_linkage() {
|
|
|
273 if (cuda_library_name != NULL) {
|
|
|
274 void *handle = dlopen(cuda_library_name, RTLD_LAZY);
|
|
|
275 if (handle != NULL) {
|
|
|
276 _cuda_cu_init =
|
|
|
277 CAST_TO_FN_PTR(cuda_cu_init_func_t, dlsym(handle, "cuInit"));
|
|
|
278 _cuda_cu_ctx_create =
|
|
|
279 CAST_TO_FN_PTR(cuda_cu_ctx_create_func_t, dlsym(handle, "cuCtxCreate"));
|
|
|
280 _cuda_cu_ctx_detach =
|
|
|
281 CAST_TO_FN_PTR(cuda_cu_ctx_detach_func_t, dlsym(handle, "cuCtxDetach"));
|
|
|
282 _cuda_cu_ctx_synchronize =
|
|
|
283 CAST_TO_FN_PTR(cuda_cu_ctx_synchronize_func_t, dlsym(handle, "cuCtxSynchronize"));
|
|
|
284 _cuda_cu_device_get_count =
|
|
|
285 CAST_TO_FN_PTR(cuda_cu_device_get_count_func_t, dlsym(handle, "cuDeviceGetCount"));
|
|
|
286 _cuda_cu_device_get_name =
|
|
|
287 CAST_TO_FN_PTR(cuda_cu_device_get_name_func_t, dlsym(handle, "cuDeviceGetName"));
|
|
|
288 _cuda_cu_device_get =
|
|
|
289 CAST_TO_FN_PTR(cuda_cu_device_get_func_t, dlsym(handle, "cuDeviceGet"));
|
|
|
290 _cuda_cu_device_compute_capability =
|
|
|
291 CAST_TO_FN_PTR(cuda_cu_device_compute_capability_func_t, dlsym(handle, "cuDeviceComputeCapability"));
|
|
|
292 _cuda_cu_device_get_attribute =
|
|
|
293 CAST_TO_FN_PTR(cuda_cu_device_get_attribute_func_t, dlsym(handle, "cuDeviceGetAttribute"));
|
|
|
294 _cuda_cu_module_get_function =
|
|
|
295 CAST_TO_FN_PTR(cuda_cu_module_get_function_func_t, dlsym(handle, "cuModuleGetFunction"));
|
|
|
296 _cuda_cu_module_load_data_ex =
|
|
|
297 CAST_TO_FN_PTR(cuda_cu_module_load_data_ex_func_t, dlsym(handle, "cuModuleLoadDataEx"));
|
|
|
298 _cuda_cu_launch_kernel =
|
|
|
299 CAST_TO_FN_PTR(cuda_cu_launch_kernel_func_t, dlsym(handle, "cuLaunchKernel"));
|
|
|
300 if (TraceGPUInteraction) {
|
|
|
301 tty->print_cr("[CUDA] Success: library linkage");
|
|
|
302 }
|
|
|
303 return true;
|
|
|
304 } else {
|
|
|
305 // Unable to dlopen libcuda
|
|
|
306 tty->print_cr("Use LD_LIBRARY_PATH (or other means) to specify installed location of CUDA library");
|
|
|
307 return false;
|
|
|
308 }
|
|
|
309 } else {
|
|
|
310 tty->print_cr("Unsupported CUDA platform");
|
|
|
311 return false;
|
|
|
312 }
|
|
|
313 tty->print_cr("Failed to find CUDA linkage");
|
|
|
314 return false;
|
|
|
315 }
|
|
|
316
|
