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 CUctxCreateParams ctxCreateParams = {};
86 CUDA_CHECK(cuCtxCreate_v4(&context, &ctxCreateParams, 0, device), "cuCtxCreate");
87 std::cout << "CudaBackend context created ok (id=" << context << ")" << std::endl;
88 dynamic_cast<CudaQueue *>(queue)->init();
89 } else {
90 CUDA_CHECK(initStatus, "cuInit() failed we seem to have the runtime library but no device");
91 }
92 }
93
94 CudaBackend::~CudaBackend() {
95 std::cout << "freeing context" << std::endl;
96 CUDA_CHECK(cuCtxDestroy(context), "cuCtxDestroy");
97 }
98
99 void CudaBackend::info() {
100 char name[100];
101 CUDA_CHECK(cuDeviceGetName(name, sizeof(name), device), "cuDeviceGetName");
102
103 std::cout << "> Using device 0: " << name << std::endl;
104
105 // get compute capabilities and the device name
106 int major = 0, minor = 0;
107 CUDA_CHECK(cuDeviceGetAttribute(&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, device), "cuDeviceGetAttribute");
108 CUDA_CHECK(cuDeviceGetAttribute(&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, device), "cuDeviceGetAttribute");
109 std::cout << "> GPU Device has major=" << major << " minor=" << minor << " compute capability" << std::endl;
110
111 int warpSize;
112 CUDA_CHECK(cuDeviceGetAttribute(&warpSize, CU_DEVICE_ATTRIBUTE_WARP_SIZE, device), "cuDeviceGetAttribute");
113 std::cout << "> GPU Device has warpSize " << warpSize << std::endl;
114
115 int threadsPerBlock;
116 CUDA_CHECK(cuDeviceGetAttribute(&threadsPerBlock, CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK, device), "cuDeviceGetAttribute");
117 std::cout << "> GPU Device has threadsPerBlock " << threadsPerBlock << std::endl;
118
119 int cores;
120 CUDA_CHECK(cuDeviceGetAttribute(&cores, CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, device), "cuDeviceGetAttribute");
121 std::cout << "> GPU Cores " << cores << std::endl;
122
123 size_t totalGlobalMem;
124 CUDA_CHECK(cuDeviceTotalMem(&totalGlobalMem, device), "cuDeviceTotalMem");
125 std::cout << " Total amount of global memory: " << (unsigned long long) totalGlobalMem << std::endl;
126 std::cout << " 64-bit Memory Address: " <<
127 ((totalGlobalMem > static_cast<unsigned long long>(4) * 1024 * 1024 * 1024L) ? "YES" : "NO") << std::endl;
128 }
129
130 PtxSource *CudaBackend::nvcc(const CudaSource *cudaSource) {
131 const uint64_t time = timeSinceEpochMillisec();
132 const std::string ptxPath = tmpFileName(time, ".ptx");
133 const std::string cudaPath = tmpFileName(time, ".cu");
134 int pid;
135 cudaSource->write(cudaPath);
136 if ((pid = fork()) == 0) { //child
137 const auto path = "/usr/local/cuda/bin/nvcc";
138 const char *argv[] {
139 "/usr/local/cuda/bin/nvcc",
140 "-ptx",
141 "-Wno-deprecated-gpu-targets",
142 cudaPath.c_str(),
143 "-o",
144 ptxPath.c_str(),
145 nullptr
146 };
147 const int stat = execvp(path, (char *const *) argv);
148 std::cerr << " nvcc stat = " << stat << " errno=" << errno << " '" << std::strerror(errno) << "'" << std::endl;
149 std::exit(errno);
150 } else if (pid < 0) {// fork failed.
151 std::cerr << "fork of nvcc failed" << std::endl;
152 std::exit(1);
153 } else { //parent
154 int status;
155 pid_t result = wait(&status);
156 auto *ptx = new PtxSource();
157 ptx->read(ptxPath);
158 return ptx;
159 }
160 }
161
162 CudaBackend::CudaModule *CudaBackend::compile(const CudaSource &cudaSource) {
163 return compile(&cudaSource);
164 }
165
166 CudaBackend::CudaModule *CudaBackend::compile(const CudaSource *cudaSource) {
167 const PtxSource *ptxSource = nvcc(cudaSource);
168 return compile(ptxSource);
169 }
170
171 CudaBackend::CudaModule *CudaBackend::compile(const PtxSource &ptxSource) {
172 return compile(&ptxSource);
173 }
174
175 CudaBackend::CudaModule *CudaBackend::compile(const PtxSource *ptx) {
176 CUmodule module;
177 if (ptx->text != nullptr) {
178 const Log *infLog = new Log(8192);
179 const Log *errLog = new Log(8192);
180 constexpr unsigned int optc = 5;
181 const auto jitOptions = new CUjit_option[optc];
182 auto jitOptVals = new void *[optc];
183
184 jitOptions[0] = CU_JIT_INFO_LOG_BUFFER_SIZE_BYTES;
185 jitOptVals[0] = reinterpret_cast<void *>(infLog->len);
186 jitOptions[1] = CU_JIT_INFO_LOG_BUFFER;
187 jitOptVals[1] = infLog->text;
188 jitOptions[2] = CU_JIT_ERROR_LOG_BUFFER_SIZE_BYTES;
189 jitOptVals[2] = reinterpret_cast<void *>(errLog->len);
190 jitOptions[3] = CU_JIT_ERROR_LOG_BUFFER;
191 jitOptVals[3] = errLog->text;
192 jitOptions[4] = CU_JIT_GENERATE_LINE_INFO;
193 jitOptVals[4] = reinterpret_cast<void *>(1);
194
195 CUDA_CHECK(cuModuleLoadDataEx(&module, ptx->text, optc, jitOptions, (void **) jitOptVals), "cuModuleLoadDataEx");
196
197 if (*infLog->text!='\0'){
198 std::cout << "> PTX JIT inflog:" << std::endl << infLog->text << std::endl;
199 }
200 if (*errLog->text!='\0'){
201 std::cout << "> PTX JIT errlog:" << std::endl << errLog->text << std::endl;
202 }
203 return new CudaModule(this, ptx->text, infLog->text, true, module);
204
205 //delete ptx;
206 } else {
207 std::cout << "no ptx content!" << std::endl;
208 exit(1);
209 }
210 }
211
212 //Entry point from HAT. We use the config PTX bit to determine which Source type
213
214 Backend::CompilationUnit *CudaBackend::compile(const int len, char *source) {
215 if (config->traceCalls) {
216 std::cout << "inside compileProgram" << std::endl;
217 }
218
219 if (config->ptx){
220 if (config->trace) {
221 std::cout << "compiling from provided ptx " << std::endl;
222 }
223 PtxSource ptxSource(len, source, false);
224 return compile(ptxSource);
225 }else{
226 if (config->trace) {
227 std::cout << "compiling from provided cuda " << std::endl;
228 }
229 CudaSource cudaSource(len , source, false);
230 return compile(cudaSource);
231 }
232 }
233
234 /*
235
236 if (config->ptx) {
237
238 } else {
239 if (config->trace) {
240 std::cout << "compiling from cuda c99 " << std::endl;
241 }
242 if (config->showCode) {
243 std::cout << "cuda " << source << std::endl;
244 }
245 auto* cuda = new CudaSource(len, source, false);
246 ptx = nvcc(cuda);
247 }
248 if (config->showCode) {
249 std::cout << "ptx " << ptx->text << std::endl;
250 }
251 CUmodule module;
252
253
254 if (ptx->text != nullptr) {
255 constexpr unsigned int jitNumOptions = 2;
256 const auto jitOptions = new CUjit_option[jitNumOptions];
257 const auto jitOptVals = new void *[jitNumOptions];
258
259 // set up size of compilation log buffer
260 jitOptions[0] = CU_JIT_INFO_LOG_BUFFER_SIZE_BYTES;
261 constexpr int jitLogBufferSize = 8192;
262 jitOptVals[0] = reinterpret_cast<void *>(jitLogBufferSize);
263
264 // set up pointer to the compilation log buffer
265 jitOptions[1] = CU_JIT_INFO_LOG_BUFFER;
266 auto jitLogBuffer = new char[jitLogBufferSize];
267 jitOptVals[1] = jitLogBuffer;
268 cuCtxSetCurrent(context);
269
270 WHERE{
271 .f = __FILE__, .l = __LINE__,
272 .e = cuModuleLoadDataEx(&module, ptx->text, jitNumOptions, jitOptions, jitOptVals),
273 .t = "cuModuleLoadDataEx"
274 }.report();
275 if (jitLogBuffer != nullptr && *jitLogBuffer!='\0'){
276 std::cout << "PTX log:" << jitLogBuffer << std::endl;
277 }
278 return new CudaModule(this, ptx->text, jitLogBuffer, true, module);
279 } else {
280 std::cout << "no ptx content!" << std::endl;
281 exit(1);
282 }
283 } */
284
285 extern "C" long getBackend(int mode) {
286 long backendHandle = reinterpret_cast<long>(new CudaBackend(mode));
287 // std::cout << "getBackend() -> backendHandle=" << std::hex << backendHandle << std::dec << std::endl;
288 return backendHandle;
289 }
290
291 void clCallback(void *) {
292 std::cerr << "start of compute" << std::endl;
293 }
294
295 void CudaBackend::computeEnd() {
296 queue->computeEnd();
297 }
298
299 void CudaBackend::computeStart() {
300 queue->computeStart();
301 }
302
303 bool CudaBackend::getBufferFromDeviceIfDirty(void *memorySegment, long memorySegmentLength) {
304 if (config->traceCalls) {
305 std::cout << "getBufferFromDeviceIfDirty(" << std::hex << reinterpret_cast<long>(memorySegment) << "," <<
306 std::dec << memorySegmentLength << "){" << std::endl;
307 }
308 if (config->minimizeCopies) {
309 const BufferState *bufferState = BufferState::of(memorySegment, memorySegmentLength);
310 if (bufferState->state == BufferState::DEVICE_OWNED) {
311 queue->copyFromDevice(static_cast<Backend::Buffer *>(bufferState->vendorPtr));
312 if (config->traceEnqueues | config->traceCopies) {
313 std::cout << "copying buffer from device (from java access) " << std::endl;
314 }
315 queue->wait();
316 queue->release();
317 } else {
318 std::cout << "HOW DID WE GET HERE 1 attempting to get buffer but buffer is not device dirty" << std::endl;
319 std::exit(1);
320 }
321 } else {
322 std::cerr <<
323 "HOW DID WE GET HERE ? java side should avoid calling getBufferFromDeviceIfDirty as we are not minimising buffers!"
324 << std::endl;
325 std::exit(1);
326 }
327 if (config->traceCalls) {
328 std::cout << "}getBufferFromDeviceIfDirty()" << std::endl;
329 }
330 return true;
331 }
332
333 CudaBackend *CudaBackend::of(const long backendHandle) {
334 return reinterpret_cast<CudaBackend *>(backendHandle);
335 }
336
337 CudaBackend *CudaBackend::of(Backend *backend) {
338 return dynamic_cast<CudaBackend *>(backend);
339 }
340
341 CudaBackend::CudaBuffer *CudaBackend::getOrCreateBuffer(BufferState *bufferState) {
342 CudaBuffer *cudaBuffer = nullptr;
343 if (bufferState->vendorPtr == nullptr || bufferState->state == BufferState::NEW_STATE) {
344 cudaBuffer = new CudaBuffer(this, bufferState);
345 if (config->trace) {
346 std::cout << "We allocated arg buffer " << std::endl;
347 }
348 } else {
349 if (config->trace) {
350 std::cout << "Were reusing buffer buffer " << std::endl;
351 }
352 cudaBuffer = static_cast<CudaBuffer *>(bufferState->vendorPtr);
353 }
354 return cudaBuffer;
355 }