1 /*
2 * Copyright (c) 2024, 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. Oracle designates this
8 * particular file as subject to the "Classpath" exception as provided
9 * by Oracle in the LICENSE file that accompanied this code.
10 *
11 * This code is distributed in the hope that it will be useful, but WITHOUT
12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 * version 2 for more details (a copy is included in the LICENSE file that
15 * accompanied this code).
16 *
17 * You should have received a copy of the GNU General Public License version
18 * 2 along with this work; if not, write to the Free Software Foundation,
19 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
20 *
21 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
22 * or visit www.oracle.com if you need additional information or have any
23 * questions.
24 */
25
26 #include <sys/wait.h>
27 #include <chrono>
28 #include "cuda_backend.h"
29 #include <iostream>
30
31 PtxSource::PtxSource()
32 : Text(0L) {
33 }
34
35 PtxSource::PtxSource(size_t len)
36 : Text(len) {
37 }
38
39 PtxSource::PtxSource(char *text)
40 : Text(text, false) {
41 }
42
43 PtxSource::PtxSource(size_t len, char *text)
44 : Text(len, text, true) {
45 }
46 PtxSource::PtxSource(size_t len, char *text, bool isCopy)
47 : Text(len, text, isCopy) {
48 }
49
50 CudaSource::CudaSource(size_t len)
51 : Text(len) {
52 }
53
54 CudaSource::CudaSource(char *text)
55 : Text(text, false) {
56 }
57
58 CudaSource::CudaSource(size_t len, char *text, bool isCopy)
59 : Text(len, text, isCopy) {
60 }
61
62 CudaSource::CudaSource()
63 : Text(0) {
64 }
65
66 uint64_t timeSinceEpochMillisec() {
67 using namespace std::chrono;
68 return duration_cast<milliseconds>(system_clock::now().time_since_epoch()).count();
69 }
70
71 std::string tmpFileName(uint64_t time, const std::string &suffix) {
72 std::stringstream timestamp;
73 timestamp << "./tmp" << time << suffix;
74 return timestamp.str();
75 }
76
77 CudaBackend::CudaBackend(int configBits)
78 : Backend(new Config(configBits), new CudaQueue(this)), initStatus(cuInit(0)), device(), context() {
79 int deviceCount = 0;
80
81 if (initStatus == CUDA_SUCCESS) {
82 CUDA_CHECK(cuDeviceGetCount(&deviceCount), "cuDeviceGetCount");
83 std::cout << "CudaBackend device count = " << deviceCount << std::endl;
84 CUDA_CHECK(cuDeviceGet(&device, 0), "cuDeviceGet");
85 CUDA_CHECK(cuCtxCreate(&context, 0, device), "cuCtxCreate");
86 std::cout << "CudaBackend context created ok (id=" << context << ")" << std::endl;
87 dynamic_cast<CudaQueue *>(queue)->init();
88 } else {
89 CUDA_CHECK(initStatus, "cuInit() failed we seem to have the runtime library but no device");
90 }
91 }
92
93 CudaBackend::~CudaBackend() {
94 std::cout << "freeing context" << std::endl;
95 CUDA_CHECK(cuCtxDestroy(context), "cuCtxDestroy");
96 }
97
98 void CudaBackend::info() {
99 char name[100];
100 CUDA_CHECK(cuDeviceGetName(name, sizeof(name), device), "cuDeviceGetName");
101
102 std::cout << "> Using device 0: " << name << std::endl;
103
104 // get compute capabilities and the device name
105 int major = 0, minor = 0;
106 CUDA_CHECK(cuDeviceGetAttribute(&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, device), "cuDeviceGetAttribute");
107 CUDA_CHECK(cuDeviceGetAttribute(&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, device), "cuDeviceGetAttribute");
108 std::cout << "> GPU Device has major=" << major << " minor=" << minor << " compute capability" << std::endl;
109
110 int warpSize;
111 CUDA_CHECK(cuDeviceGetAttribute(&warpSize, CU_DEVICE_ATTRIBUTE_WARP_SIZE, device), "cuDeviceGetAttribute");
112 std::cout << "> GPU Device has warpSize " << warpSize << std::endl;
113
114 int threadsPerBlock;
115 CUDA_CHECK(cuDeviceGetAttribute(&threadsPerBlock, CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK, device), "cuDeviceGetAttribute");
116 std::cout << "> GPU Device has threadsPerBlock " << threadsPerBlock << std::endl;
117
118 int cores;
119 CUDA_CHECK(cuDeviceGetAttribute(&cores, CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, device), "cuDeviceGetAttribute");
120 std::cout << "> GPU Cores " << cores << std::endl;
121
122 size_t totalGlobalMem;
123 CUDA_CHECK(cuDeviceTotalMem(&totalGlobalMem, device), "cuDeviceTotalMem");
124 std::cout << " Total amount of global memory: " << (unsigned long long) totalGlobalMem << std::endl;
125 std::cout << " 64-bit Memory Address: " <<
126 ((totalGlobalMem > static_cast<unsigned long long>(4) * 1024 * 1024 * 1024L) ? "YES" : "NO") << std::endl;
127 }
128
129 PtxSource *CudaBackend::nvcc(const CudaSource *cudaSource) {
130 const uint64_t time = timeSinceEpochMillisec();
131 const std::string ptxPath = tmpFileName(time, ".ptx");
132 const std::string cudaPath = tmpFileName(time, ".cu");
133 int pid;
134 cudaSource->write(cudaPath);
135 if ((pid = fork()) == 0) { //child
136 const auto path = "/usr/local/cuda/bin/nvcc";
137 const char *argv[] {
138 "/usr/local/cuda/bin/nvcc",
139 "-ptx",
140 "-Wno-deprecated-gpu-targets",
141 cudaPath.c_str(),
142 "-o",
143 ptxPath.c_str(),
144 nullptr
145 };
146 const int stat = execvp(path, (char *const *) argv);
147 std::cerr << " nvcc stat = " << stat << " errno=" << errno << " '" << std::strerror(errno) << "'" << std::endl;
148 std::exit(errno);
149 } else if (pid < 0) {// fork failed.
150 std::cerr << "fork of nvcc failed" << std::endl;
151 std::exit(1);
152 } else { //parent
153 int status;
154 pid_t result = wait(&status);
155 auto *ptx = new PtxSource();
156 ptx->read(ptxPath);
157 return ptx;
158 }
159 }
160
161 CudaBackend::CudaModule *CudaBackend::compile(const CudaSource &cudaSource) {
162 return compile(&cudaSource);
163 }
164
165 CudaBackend::CudaModule *CudaBackend::compile(const CudaSource *cudaSource) {
166 const PtxSource *ptxSource = nvcc(cudaSource);
167 return compile(ptxSource);
168 }
169
170 CudaBackend::CudaModule *CudaBackend::compile(const PtxSource &ptxSource) {
171 return compile(&ptxSource);
172 }
173
174 CudaBackend::CudaModule *CudaBackend::compile(const PtxSource *ptx) {
175 CUmodule module;
176 if (ptx->text != nullptr) {
177 const Log *infLog = new Log(8192);
178 const Log *errLog = new Log(8192);
179 constexpr unsigned int optc = 5;
180 const auto jitOptions = new CUjit_option[optc];
181 auto jitOptVals = new void *[optc];
182
183 jitOptions[0] = CU_JIT_INFO_LOG_BUFFER_SIZE_BYTES;
184 jitOptVals[0] = reinterpret_cast<void *>(infLog->len);
185 jitOptions[1] = CU_JIT_INFO_LOG_BUFFER;
186 jitOptVals[1] = infLog->text;
187 jitOptions[2] = CU_JIT_ERROR_LOG_BUFFER_SIZE_BYTES;
188 jitOptVals[2] = reinterpret_cast<void *>(errLog->len);
189 jitOptions[3] = CU_JIT_ERROR_LOG_BUFFER;
190 jitOptVals[3] = errLog->text;
191 jitOptions[4] = CU_JIT_GENERATE_LINE_INFO;
192 jitOptVals[4] = reinterpret_cast<void *>(1);
193
194 CUDA_CHECK(cuModuleLoadDataEx(&module, ptx->text, optc, jitOptions, (void **) jitOptVals), "cuModuleLoadDataEx");
195
196 if (*infLog->text!='\0'){
197 std::cout << "> PTX JIT inflog:" << std::endl << infLog->text << std::endl;
198 }
199 if (*errLog->text!='\0'){
200 std::cout << "> PTX JIT errlog:" << std::endl << errLog->text << std::endl;
201 }
202 return new CudaModule(this, ptx->text, infLog->text, true, module);
203
204 //delete ptx;
205 } else {
206 std::cout << "no ptx content!" << std::endl;
207 exit(1);
208 }
209 }
210
211 //Entry point from HAT. We use the config PTX bit to determine which Source type
212
213 Backend::CompilationUnit *CudaBackend::compile(const int len, char *source) {
214 if (config->traceCalls) {
215 std::cout << "inside compileProgram" << std::endl;
216 }
217
218 if (config->ptx){
219 if (config->trace) {
220 std::cout << "compiling from provided ptx " << std::endl;
221 }
222 PtxSource ptxSource(len, source, false);
223 return compile(ptxSource);
224 }else{
225 if (config->trace) {
226 std::cout << "compiling from provided cuda " << std::endl;
227 }
228 CudaSource cudaSource(len , source, false);
229 return compile(cudaSource);
230 }
231 }
232
233 /*
234
235 if (config->ptx) {
236
237 } else {
238 if (config->trace) {
239 std::cout << "compiling from cuda c99 " << std::endl;
240 }
241 if (config->showCode) {
242 std::cout << "cuda " << source << std::endl;
243 }
244 auto* cuda = new CudaSource(len, source, false);
245 ptx = nvcc(cuda);
246 }
247 if (config->showCode) {
248 std::cout << "ptx " << ptx->text << std::endl;
249 }
250 CUmodule module;
251
252
253 if (ptx->text != nullptr) {
254 constexpr unsigned int jitNumOptions = 2;
255 const auto jitOptions = new CUjit_option[jitNumOptions];
256 const auto jitOptVals = new void *[jitNumOptions];
257
258 // set up size of compilation log buffer
259 jitOptions[0] = CU_JIT_INFO_LOG_BUFFER_SIZE_BYTES;
260 constexpr int jitLogBufferSize = 8192;
261 jitOptVals[0] = reinterpret_cast<void *>(jitLogBufferSize);
262
263 // set up pointer to the compilation log buffer
264 jitOptions[1] = CU_JIT_INFO_LOG_BUFFER;
265 auto jitLogBuffer = new char[jitLogBufferSize];
266 jitOptVals[1] = jitLogBuffer;
267 cuCtxSetCurrent(context);
268
269 WHERE{
270 .f = __FILE__, .l = __LINE__,
271 .e = cuModuleLoadDataEx(&module, ptx->text, jitNumOptions, jitOptions, jitOptVals),
272 .t = "cuModuleLoadDataEx"
273 }.report();
274 if (jitLogBuffer != nullptr && *jitLogBuffer!='\0'){
275 std::cout << "PTX log:" << jitLogBuffer << std::endl;
276 }
277 return new CudaModule(this, ptx->text, jitLogBuffer, true, module);
278 } else {
279 std::cout << "no ptx content!" << std::endl;
280 exit(1);
281 }
282 } */
283
284 extern "C" long getBackend(int mode) {
285 long backendHandle = reinterpret_cast<long>(new CudaBackend(mode));
286 // std::cout << "getBackend() -> backendHandle=" << std::hex << backendHandle << std::dec << std::endl;
287 return backendHandle;
288 }
289
290 void clCallback(void *) {
291 std::cerr << "start of compute" << std::endl;
292 }
293
294 void CudaBackend::computeEnd() {
295 queue->computeEnd();
296 }
297
298 void CudaBackend::computeStart() {
299 queue->computeStart();
300 }
301
302 bool CudaBackend::getBufferFromDeviceIfDirty(void *memorySegment, long memorySegmentLength) {
303 if (config->traceCalls) {
304 std::cout << "getBufferFromDeviceIfDirty(" << std::hex << reinterpret_cast<long>(memorySegment) << "," <<
305 std::dec << memorySegmentLength << "){" << std::endl;
306 }
307 if (config->minimizeCopies) {
308 const BufferState *bufferState = BufferState::of(memorySegment, memorySegmentLength);
309 if (bufferState->state == BufferState::DEVICE_OWNED) {
310 queue->copyFromDevice(static_cast<Backend::Buffer *>(bufferState->vendorPtr));
311 if (config->traceEnqueues | config->traceCopies) {
312 std::cout << "copying buffer from device (from java access) " << std::endl;
313 }
314 queue->wait();
315 queue->release();
316 } else {
317 std::cout << "HOW DID WE GET HERE 1 attempting to get buffer but buffer is not device dirty" << std::endl;
318 std::exit(1);
319 }
320 } else {
321 std::cerr <<
322 "HOW DID WE GET HERE ? java side should avoid calling getBufferFromDeviceIfDirty as we are not minimising buffers!"
323 << std::endl;
324 std::exit(1);
325 }
326 if (config->traceCalls) {
327 std::cout << "}getBufferFromDeviceIfDirty()" << std::endl;
328 }
329 return true;
330 }
331
332 CudaBackend *CudaBackend::of(const long backendHandle) {
333 return reinterpret_cast<CudaBackend *>(backendHandle);
334 }
335
336 CudaBackend *CudaBackend::of(Backend *backend) {
337 return dynamic_cast<CudaBackend *>(backend);
338 }
339
340 CudaBackend::CudaBuffer *CudaBackend::getOrCreateBuffer(BufferState *bufferState) {
341 CudaBuffer *cudaBuffer = nullptr;
342 if (bufferState->vendorPtr == nullptr || bufferState->state == BufferState::NEW_STATE) {
343 cudaBuffer = new CudaBuffer(this, bufferState);
344 if (config->trace) {
345 std::cout << "We allocated arg buffer " << std::endl;
346 }
347 } else {
348 if (config->trace) {
349 std::cout << "Were reusing buffer buffer " << std::endl;
350 }
351 cudaBuffer = static_cast<CudaBuffer *>(bufferState->vendorPtr);
352 }
353 return cudaBuffer;
354 }